ComfyUI Detailer/ADetailer Workflow
If you're having ANY issues with this workflow, please include the error log from your console when writing a comment here on this page. Chances are 99.98% i can help you with your issue - but i need the error log. And me not having to ask for it first will save you potentially hours. Chances are also high the solution to your problem is already documented in the FAQ here.
You can now support me on my Ko-Fi if you have too much money (✿◡‿◡)
Requirements for each version are listed below or can be found inside a Note in the Workflow itself.
Because of the many connections among the nodes, I highly recommend turning off the link visibility by clicking the "Toggle Link visibility" (Eye icon) in the bottom right of ComfyUI.
Description
I originally came from A1111 WebUI to ComfyUI and, honestly, a lot of things felt way more complicated than they needed to be. I couldn’t find workflows that were both visually pleasing and showed me all the important options without requiring a deep understanding of every little thing in ComfyUI.
Over a few months I learned the ins and outs of Comfy and built my own very barebones, simple workflow (v1) with one main goal: streamline the process and make it visually understandable. I released it here thinking it might help others who also want to make the jump from A1111 to Comfy.
Over the last year I’ve kept adding more and more features as people started using the workflow and requesting things. The workflow has grown a lot, but I’m still trying to keep the same ease-of-use that this whole journey started with. The main goal is still the same: give you a lot of powerful options in a layout that’s as clean, readable, and as user-friendly as ComfyUI will allow.
At this point I feel like I have a pretty solid understanding of how Comfy works, and I’ve even created my own custom nodes to add missing functionality so this workflow could match the ideas I had for it. I try to hide as much of the technical complexity as possible — but if you’re ever curious or confused about anything, please feel free to ask!
Thanks to all of your suggestions, the workflow now includes features like:
Single-image and batch generation
Automatic detailers for specific body parts
Upscaling
v-pred models
LoRAs
ControlNet
IPAdapter
Hires fix / refiner
Manual inpainting
Thank you to every single person who uses this workflow, donates Buzz, or shares images on this page.
I really appreciate you taking those extra steps to support and promote the project. ♥️
Requirements
v4.4 and above require ComfyUI version 0.3.51 or above and also need the frontend to be AT LEAST 1.24.3 or later.
v5.1 - Full List
v5.0 - Full List
v4.4 - Additions to v4.3 (IMG2IMG Only)
Otherwise same as v4-4.3 below
v4.2-4.3 - Additions to v4.1
Otherwise same as v4.1 (incl. v4) below
v4.1 - Additions to v4 (IMG2IMG Only)
Otherwise same as v4 below
v4 - Full List
v3 - v3.2 - Full List
v2.2 - Additions to v2
Otherwise same Custom-Nodes as v2 but you can remove Comfyui-ergouzi-Nodes
v2 - Full List
v1 - Full List
How to use
Since all of the different versions work differently, you should check the "How to use" Node inside of the Workflow itself.
I promise that once you read the explanation of the workflow itself, it'll click and it will be a simple plug and play experience.
It's the simplest I could've made it coming from someone who's only started using ComfyUI 4-5 months ago and had been exclusively an A1111WebUI user before.
When were what functionalitys added?
Starting from v3, ControlNet is included.
Starting from v4, IPAdapter is included.
Starting from v4.3, HiRes Fix and Dynamic prompts(wildcards) is included.
Starting from v4.4, Refiner is included.
Starting from v5.0, Manual Inpainting is included.
Any errors during execution?
If you're running into any errors during the execution of the workflow, please check the FAQ of my Guide first. The guide is written for the IMG2IMG Workflow but when issues arise that people run into frequently i'll add the solutions and what's hapenning to that FAQ section.
If you can't find the problem you're running into there - feel free to write me a comment on the model page so that i can help you and other people might benefit from it as well.
Feedback
I'd love to see your feedback or opinion on the workflow.
This is the first workflow I have ever created myself from scratch and I'd love to hear what you think of it.
If you want to do me a huge favor, you can post your results on this Model page.
I'll make sure to send some buzz your way!
Description
(bugfix): changed color fix algorithm and strength to fix noise in dark areas
this is a minor bugfix release, no need to download if you don't have any issues with hires fix and the color fix
FAQ
Comments (150)
Hello, I have wanted to try using ControlNet and am running into an error.
I read the guide in the template. and it seems like you have the option to choose either using the 'Pre-Process image to pose' node, or the 'Load ControlNetPose' node. I put an image into the Pre-process image to pose' node but am receiving a red outline on the controlnetpose node. even though im not trying to use it. I have both the enable controlnet and enable preprocessor switches to yes.
inside the console i get a "- Custom validation failed for node: image - Invalid image file: 0ac489a4b8cc52fcfc47f2c42a40bfeab694c657700a7bea4d66b914acee9658.jpg" which explains the red error outline but again i thought i had the choice between which node i wanted to use. the image file in the log is just what it was set to by defualt when i loaded the template.
apologies if im being unclear, im a bit dyslexic lul
i can try and reword if it's confusing.
Heya @Krow_Taxes,
this happens because with the new comfyui versions all inputs of active nodes are being validated for a valid input.
I don't think you have downloaded the workflow from here then, maybe you have downloaded it through an example image of someone else or have switched the input of that image with something you previously had in your input folder?
I do not have any files that are named like this and before releasing any of my workflows i always switch all image inputs to the "example.png" because that is present automatically in every comfyui installation.
So you do need to have an image selected in the controlnet node, using it with the preprocessing just switches the input for controlnet to the preprocessing image.
I just recommend switching the input of the node back to the example.png or re-downloading the workflow. The example.png is always in your comfyui/input folder - otherwise you can also just load any random image you want since it won't be used. You can also use the arrow buttons next to the image name to switch between them or open a dropdown to select a filename from your input folder.
@vslinx you are right about that image name being one of my own that i deleted, that was a misspeak.
Here's a question, in the ControlNet Preview node if i had everything set up correctly, would it be showing the preprocessed output in the preview? or would it still just be showing the controlnet pose image?
If it is the latter, is there some way i can see the pre-processed output just to make sure it's working? because i've tried looking through the layout myself and all i can get is the same image as the pre-processed-image-to-pose input image i set without any edits.
Again, apologies if I'm misunderstanding how ControlNet is supposed to work.
@Krow_Taxes when you have pre-processing enabled as well as controlnet, the controlnet preview should always display the pre-processed image as a pose depending on the pre-processing model you've selected in the "ControlNet Models"-Node.
Make sure you have your comfyui and custom nodes up-to-date and have the current version of the workflow (v4.5) as 4.3 for example will have compatibility issues with the new comfyui versions due to recent changes in comfy. Maybe re-download the workflow if you've already made changes to make sure you have the default workflow.
I have tested it just now and with pre-processing enabled the input image of the pre-process image node will be turned into a pose and shown in the preview pose node in the debug section.
Hello! i have this error whenever i try to use detection models:
UltralyticsDetectorProvider
Weights only load failed. This file can still be loaded, to do so you have two options, [1mdo those steps only if you trust the source of the checkpoint[0m. (1) In PyTorch 2.6, we changed the default value of the weights_only argument in torch.load from False to True. Re-running torch.load with weights_only set to False will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source. (2) Alternatively, to load with weights_only=True please check the recommended steps in the following error message. WeightsUnpickler error: Unsupported global: GLOBAL ultralytics.nn.tasks.SegmentationModel was not an allowed global by default. Please use torch.serialization.add_safe_globals([ultralytics.nn.tasks.SegmentationModel]) or the torch.serialization.safe_globals([ultralytics.nn.tasks.SegmentationModel]) context manager to allowlist this global if you trust this class/function. Check the documentation of torch.load to learn more about types accepted by default with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.
Heya @Desmond93,
loading model files can involve executing code, so if malicious code is embedded in a model, it can pose a security risk. For this reason, PyTorch 2.6 and later have introduced features for safer model loading. The issue is that older model files, which were created without such safe loading (weights_only) restrictions, may occasionally fail to load properly. So there are 3 things you can do to fix this.
1. You can make sure ComfyUI-Impact-Pack and ComfyUI-ImpactSubpack are up-to-date and try again. If that doesn't work the older version of your ultralytics model is not compatible with the weights_only method of your pytorch version. In that case you can either do #2 or #3(if you trust the model)
2. You can update your PyTorch version, i have not had any issues so far with any ultralytics model on PyTorch 2.8.0. To do this first check what version you have - when starting comfyui in the console you should see something like this: "pytorch version: 2.8.0+cu128".
If you already have 2.8.0 and it doesn't work you can only do option #3, otherwise i would recommend updating your pytorch version to a newer one. You can do that by either installing a new comfyui instance or by updating the dependency. If you have the normal comfy you can update by going into your "ComfyUI\venv\Scripts"-Folder and opening a Terminal there and then executing "call activate.bat" followed by (without the " of course):
"pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128"
If you have portable comfy on your computer you can do this by going into the main folder with a terminal (where run_nvidia_gpu.bat is located) and executing this:
"python_embeded\python.exe -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128"
Lastly you need to upgrade your cuda toolkit of your GPU with this:
https://developer.nvidia.com/cuda-12-8-0-download-archive
3. If you don't want to update your pytorch or re-install comfyui AND you trust the model (important!) you can also simply add it to your whitelist. For that go into:
"ComfyUI\user\default\ComfyUI-Impact-Subpack" and you'll find a model-whitelist.txt
Here you can add one model per line that you want to permit to be loaded regardless of their old potentially malicious architecture. Don't forget to add the subfolder in which it is in, example format would be :
bbox/face_old.pt
bbox/hands_old.pt
segm/maskdetailer.pt
You can also find info's for that in the impact-subpack repo here.
I hope this helps :D
@vslinx Method #1 worked! I reinstalled ComfyUI-Impact-Pack and ComfyUI-ImpactSubpack, and now it works!
Thank you very much! I spent an entire evening using chat-gpt and couldn't solve it, but it only took 1 minute to fix!
@Desmond93 nice! i'm glad it was that easy to fix 🥰
Sadly most LLM's like ChatGPT will always try to make YOU fix the issue from the code instead of understanding where the issue comes from and analyzing from there on out ^^
So never be shy to ask if you're running into any issues with the workflow, i'll most likely have seen most of the errors already :D
Enjoy generating with the workflow! 🫡♥️
@vslinx With the last update of ComfyUI, whenever i use a Detailer it give me this error:
0: 640x640 1 face, 6.7ms
Speed: 1.9ms preprocess, 6.7ms inference, 1.1ms postprocess per image at shape (1, 3, 640, 640)
!!! Exception during processing !!! 'DifferentialDiffusion' object has no attribute 'apply'
Traceback (most recent call last):
File "D:\ComfyUI_windows_portable\ComfyUI\execution.py", line 496, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_portable\ComfyUI\execution.py", line 315, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_portable\ComfyUI\execution.py", line 289, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "D:\ComfyUI_windows_portable\ComfyUI\execution.py", line 277, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "D:\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-impact-pack\modules\impact\impact_pack.py", line 876, in doit
enhanced_img, cropped_enhanced, cropped_enhanced_alpha, mask, cnet_pil_list = FaceDetailer.enhance_face(
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-impact-pack\modules\impact\impact_pack.py", line 830, in enhance_face
DetailerForEach.do_detail(image, segs, model, clip, vae, guide_size, guide_size_for_bbox, max_size, seed, steps, cfg,
File "D:\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-impact-pack\modules\impact\impact_pack.py", line 304, in do_detail
model = nodes_differential_diffusion.DifferentialDiffusion().apply(model)[0]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'DifferentialDiffusion' object has no attribute 'apply'
@Desmond93 thanks for the headsup!
it seems like this has been announced by comfyui beforehand but sadly the Impact-Pack does not have the changes implemented yet. So there are a few open issues regarding this in the github repo:
https://github.com/ltdrdata/ComfyUI-Impact-Pack/issues
i'll add a hint to the beginning of the description to not update, until then we probably have to wait a day or two until ltrdata releases a fix :/
@Desmond93 if you want to continue using the detailer until the impact-pack developer fixes it, you can do this change locally:
@Desmond93 ltrdata has now fixed the issue in the official repo, so you can re-install the node now and use it with the newest version. The problem has been solved 🫡
How can I change the upscale model used?
@ToppaTengu in the second row (Model Backend) there is a "Upscale Model"-Node in which you can change the upscale model that is being used.
Woah, look at all the buttons!
@PantyMilk More buttons = better 😂
Thank you so much for the buzz! ♥️
Can you please tell me how to properly using Detailer with upscaler and hires? I want to change hands on upscaled version of genereted picture. What node i should start? Because if i just use Run command, it going to past all nodes - generate, upscalling and hands detailer
Hey @Velter,
what are you looking for exactly?
You can select everything you want to do yourself, so if you only want to change/detail the hands of an already generated image you should use the img2img workflow to load your image and then download a ultralytics model that detects hands and activate the hand detailer as well as the hand prompt and write what you want to create.
If you're not really sure about the whole process i'd recommend reading the guide for the IMG2IMG workflow, you can find a banner that leads to the guide in the description.
There's also a section explicitly explaining how to just detail an already existing image:
https://civitai.com/articles/15480#detail-features-of-image
If you want to completely change the hand and not just detail it but replace it with something entirely different i recommend waiting for the 5.0 that i am working on right now. It has a manual inpainting built-in so you can draw onto the image what you want to change.
@vslinx okay, i just think if there is detailer that it can using like a img2img, im sorry, im gonna try using your img2img workflow
@Velter you can do that, you just have to increase the denoise a little and prompt correctly what you intend to generate ^^
You can follow the example in the article and adjust the values accordingly.
Having pretty much the same issue as ChinuaTheRageBear where I get this error:
Some nodes require a newer version of ComfyUI (current: 0.3.59). Please update to use all nodes.
Requires ComfyUI 0.3.56:
752d53ed-f222-4a68-b4d6-a351a15c521a
Requires ComfyUI 0.3.51:
0941940b-e86a-427a-bd29-51cf47d05345, 2a07f710-6ca0-41f6-8510-f4e8adbdb404, 67a0508f-aa7b-4ba2-8265-d3fae6da4226, 6c2eee0d-5bd7-4029-9333-7f1b0ed688f2, 9726ad71-5451-4f61-8889-dbbd630619dc, c0330ef1-4c81-486f-99aa-b2f3e41be549, d63f9638-9bde-4613-a8dc-faa6238b912b, dd15cf75-15d9-42b1-8ee0-fd4bae8e8c33, f8d9af42-d1b5-4ef5-a349-267ed8b6186e
I am using the portable version, I updated to the latest version (0.3.59) and have frontend version 1.27.5. I've also deleted all of my custom nodes and reinstalled them all but still have the issue.
In the manager when I click install all missing custom nodes I get this error:
Failed to find the following ComfyRegistry list. The cache may be outdated, or the nodes may have been removed from ComfyRegistry.
comfy-core
comfy-core
comfy-core
comfy-core
comfy-core
comfy-core
comfy-core
comfy-core
comfy-core
comfy-core
And I get this pop-up:
Alert
ComfyUI is outdated, so some built-in nodes cannot be used.
Not sure what to do from here, so let me know if you guys know what I can try. Thanks!
Hey @tplesma,
this seems to sadly be a very common issue that people have with portable comfy.
It seems like the update process often breaks in the portable version resulting in parts of your comfy being updated (including the parts displaying the version), while other parts that custom-nodes rely on like dependencys and packages are not correctly updated.
As stated in the github issue i've linked above you have to update your comfyui (again) by manually executing the update.bat again. I've also seen in other issues that sometimes comfy updated but stayed on the old branch. So if you have git installed and checked the portable version out instead of downloading it you could try a "git checkout master" followed by a "git pull".
Other than that i can only recommend downloading the regular comfy. The only benefit the portable version has is that it comes with it's own python so you don't have to install that, but i feel like all the downsides i regularly see users have, is not really worth it :/
If you install comfy in a virtual environment none of the dependencys will be installed globally on your pc either, they'll all only be installed in your comfy instance.
If you want to give the full comfy a try you could try using my script, it'll check if you have comfy installed and if not downloads and opens the installer for you. It'll download everything required and install it inside a virtual environment so it won't bloat your pc. It'll also already download all custom nodes required for my workflow as well as my workflows (txt2img & img2img). I'm always using it to set up a quick test instance before releasing a new version.
If you want to give it a try you can find it here: https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/blob/main/windows-nvidia.bat
I already had a few users re-install their comfy through this and it has worked well for them so far. (You can also check what it does so you know there's nothing malicious in there).
Otherwise i can only recommend you to try updating portable comfy again or maybe re-downloading and just moving your models & custom_nodes folder over. Apart from this you can obviously also raise an issue in the comfyui github repository since it's already a well known issue.
Holeee shit. I finally figured out what the issue was for me - probably this is a noob mistake. I was using the desktop version and updated comfyui and the frontend, even tried the portable and did the update.bat. What the issue was I was using microsoft edge to open the client, which was more lightweight for me on memory. When I switched to firefox with the updated comfyui and frontend, no more errors!
@whitetable1234309 Wow that is the weirdest thing i‘ve ever heard! Tbf i never seen anyone use edge at all, but since it‘s also chromium based nowadays i wouldn‘t have thought there‘d be any issues :O
Have you always after updating refreshed your browser window? it‘s often not clear for a lot of people why you‘d need to do that but comfy caches so aggresively that even after removing and wiping and/or reinstalling comfy, the moment you open the comfy address again it loads all the old information. Meaning that it‘s technically still loading on old software. Thats why when updating or installing nodes through the manager it always asks you to refresh(f5) your browser.
But if you‘ve always done that and edge still caused issues that‘d be wild! :D will definitely keep that in mind if someone else ever runs into this issue.
@vslinx yep, I've always refreshed the browser after updating/restarting. I guess there are a bunch of reasons why the nodes don't properly show up. Hope this helps someone else. Thanks for the quick response!
@whitetable1234309 how did you switch to Firefox?
I have the similar issue. Nodes names look like 2a07f710-6ca0-41f6-8510-f4e8adbdb404, 6c2eee0d-5bd7-4029-9333-7f1b0ed688f2, dd15cf75-15d9-42b1-8ee0-fd4bae8e8c33, 752d53ed-f222-4a68-b4d6-a351a15c521a, 0941940b-e86a-427a-bd29-51cf47d05345, d63f9638-9bde-4613-a8dc-faa6238b912b, c0330ef1-4c81-486f-99aa-b2f3e41be549, f8d9af42-d1b5-4ef5-a349-267ed8b6186e, 9726ad71-5451-4f61-8889-dbbd630619dc, 67a0508f-aa7b-4ba2-8265-d3fae6da4226
ComfyUI 0.3.62
@hanryjean06653 try the steps i mentioned above.
If that doesn't work you can try what whitetable1234309 did and try a different browser.
I'd generally recommend not using comfy portable if that's what you're using.
@vslinx I'd tried Firefox, the situation got better, but nodes still don't work
@vslinx Should I delete the portable version of ComfyUI before installing your vesion?
@hanryjean06653 you can just install the full comfy version somewhere else.
Just remember that it executes wherever you start it, so put it into the place where you want comfy installed :)
And then let me know if the install script worked for you ^^
I love this workflow! I've had really great results so far. I just have one small worry; I have noticed that the sliders have disappeared and aren't able to be edited, making it difficult to up the denoising when necessary. I've tried the recommended refresh, updating etc--I don't know what else could be wrong. I searched around online, too, but I didn't find much in regards to it.
hey @pinkmoth068,
i don't know why civitai starts censoring words but what has disappeared? :D
Civitai is sadly only rendering "*********rs" so i don't really know what the issues is :/
Edit: now civitai is showing "sliders" wtf is going on with this platform 😂
The sliders should not have disappeared, can you send me a screenshot? Can you open the detailers and check if there are still nodes connected to the denoise for example?
It sounds like maybe something went wrong when you saved the workflow. It could also be that when you moved around the workflow you might have moved the sliders behind other nodes.
Can you download the workflow again and check if the default version still has everything?
@vslinx Hi, I'm really sorry I didn't see this! I don't know why I'm not getting notified, lol. I can send a screenshot for sure.
Hi, loving the workflow. I've always been trying to figure out how to use the "advanced" stuff like ultralytics and masks and all that, but never willing to put enough time into figuring out to do so, so this really helped me a lot.
I see that you're working on a new version where both txt2img and img2img would be in the same workflow, and I'm going to look forward to it. Might I also request if you could include a regional prompter section, where you could use different models for each area/section? I think I've read that before somewhere that it's possible, but it's another thing that I never quite figured how to do. If you could include that in a future version, that would be great.
Again, thanks for such an extensive yet simple workflow.
Hey @last27683599,
Thank you for your feedback and i'm glad that the workflow helped you to get into it a bit!
I'm currently working on the 5.0 which will have a manual inpainting area for both the img2img as well as the txt2img workflow. With that i'll have implemented every feature i had on my roadmap and then i'll work on finishing the true AIO workflow where txt2img and img2img are both in one workflow.
I have also used regional prompt workflows in the past and it's definitely a nice feature but sadly i don't think that i'll be implementing it (for now) into the workflow. Reason for that is mostly how much space it'd require for the function and how niche it is.
But with the implementation of Inpainting you'll be able to draw a mask (just like with the regional prompting) that covers a full character so you can simply use a character lora and replace the full character.
If the demand is there i might look into the feature after 5.0 and the AIO workflow are released to see how well it's implementable into the workflow itself. So maybe it will come! :D
Thanks again for the feedback and the compliments! really happy to hear you've been enjoying the workflow so far 🥰♥️
I can't download the workflow from the page. Is the link down? I've also tried using another browser, but nothing. Can someone please send me the workflow? I'm very interested in the project.
@mediimedii1medii448 there shouldn't be any problems downloading the model, but maybe civitai is having some issues right now with their providers.
You can always just download the example image (the picture of the girl) by clicking on it and then clicking on the "download"-button in the top right corner of the image and drag it into comfyui. That should work (at least until civitai has fixed their server issues).
Otherwise i also have the versions backed up on huggingface, so you can find the IMG2IMG and TXT2IMG workflows there. Just download the json there
Hi, I have a problem with image generation. It freezes at about 80% and then crashes.
Hey @bacsvord123464,
what is the console saying?
@vslinx Thank you, I was able to solve the problem myself, but I have a question regarding image generation. I have a pretty good picture, but I want to make several variations of this particular seed or something like that. How can this be done?
@bacsvord123464 if you have generated this image through the txt2img workflow and want to have slight variations of it you could load it into the "Pre-Process image to pose"-Node, activate ControlNet (after selecting your models) and generate images by using the same prompt you have used to generate your image.
If you instead just have an image you could load it into the img2img workflow and just hit generate. You can also find a img2img guide inside the description here.
How can I change the aspect ratio of the images generated? All the pictures end up being 1024 x 1536, I want to generate pictures at 1440 x 1850, where do I have to fidget to achieve that?
Thanks in advance, workflow works wonderfully otherwise.
@angelhante
If you're using the TXT2IMG workflow you can set the height & width at the "Size" Node (very first node top left).
You can either use the visual depiction to set the image size by moving the little white dot on the grid or you can just double click on the two values each and set it to whatever value you want.
The upper value is width and the lower value is height.
The IMG2IMG workflow always sets the smallest side of the image to 1024px while keeping aspect ratio since SDXL based models are performing best at a 1024px resolution.
@vslinx Oh, I was talking about IMG2IMG, I guess I'll have to change the reference images to the desired aspect ratio before processing then, thank you!
@angelhante if you're using the the IMG2IMG workflow you could also change the upscale factor to 1.4 and activate upscaling. That'd put the smallest side after upscaling to around 1432px, if you want to have exactly 1440px for the smallest side you could also go down left (Model Backend Group) and there will be a little node called "Default resolution".
Here you can change the value from 1024px to whatever you want and this will be used as the default smallest side of the image. The aspect ratio will always be kept during IMG2IMG Transfer to match the input image 🫡♥️
i love this workflow, its easy to use, perfect quality. can you add the shoes(feet) detailer, hair detailer, instantid workflow
thanks for the feedback @vincentwinner!
you can realistically use any detailer you want, if you're not using the lips or nose detailer you can simply rename them and instead use a ultralytics model trained on shoes/feet/hair etc.
InstantID is currently not on my planned feature list but i'll see if that's something i'm interested to do in the future as i'm currently still working on manual inpainting :)
Hey, I have a question, will this workflow work fine with low vram? I have 1660ti, so when I making generation on sdxl I have problems, it took maybe 1 hour to make 1 step, so I have fast workflow to this, but that workflow don't have much options.
And also a question, I need to download dependencies manually, or it is possible to download automatically?
hey @maks777coreb577 the workflow should work with low vram.
For options like controlnet and ipadapter i have implemented low vram options specifically where it's possible.
6GB is sadly very little even for SDXL but you can improve your generation times a lot when you're using speed loras. If you use the correct settings you won't really lose quality either.
To heavily increase the speed i recommend downloading this lora:
https://civitai.com/models/1608870/dmd2-speed-lora-sdxl-pony-illustrious
And using it in whatever workflow you'll decide to use. Set it to 0.7 strength, the steps to 8, sampler to lcm, scheduler to normal or simple, cfg/guidance to 1.6 - this alone should already HEAVILY increase your speed.
You can obviously also use this trick in my workflow for faster generation.
You might still have slow times if you're using hires fix, there will be an improvement in the next version i'm working on right now that'll work better on lower end devices when using speed loras.
I have created an install script that i use to test my workflows before releasing, it'll download and install comfy in the newest version + all required dependencies. So you can use this to install a fresh comfy installation if you want to:
https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/blob/main/windows-nvidia.bat
It will install everything required and ask you if you want to download additional stuff. The installation happens in the folder where you execute it, so drag+drop it where you want to install comfy. I do not have implemented yet for it to recognize existing installations and only download the requirements since i mainly use it for myself to test workflows before release and i always do that in a clean installation.
Will you ever add outpainting?
@Shuffle_V2 I'm adding Manual inpainting right now which also will allow outpainting.
I won't support it straight away and work on it rather in a 5.1 version but it'll already be technically possible. So if you have a little affinity with comfy you'll be able to already use it then.
Running into an issue with the Schedule Selector giving a return type mismatch to the Ksamplers, Is there a solution as it only has started happening today
@Batheem without seeing the error I assume you have installed the Clownshark custom node (https://github.com/ClownsharkBatwing/RES4LYF)
This node overwrites the default list of schedulers which results in issues when handing over the scheduler from one node to another. This is very very bad practice from the node creator and i recommend deactivating/uninstalling it.
Uninstalled, did a complete reinstall of Comfy + Custom nodes, seems to work just fine now. Thanks
Im trying to use lustify sdxl as a refiner, and Cyber realistic pony as main model. I believe clip skip is the issue and I have to put ClipSetLastLayer somewhere to change clips skip to 1 for sdxl, Am I on good path?
hey @sssebinho303,
i'm not really familiar with these models and usually it's not a great idea to mix architecture like SDXL with Pony - even if both are based on SDXL. Since Pony is re-trained on a very different language as well as prompt set.
I can see that CyberRealistic Pony does use Clip Skip 2, there is no mention of what Clip Skip setting Lustify uses. If you want to find out i'd recommend not using the refiner, generating 1 image with CyberRealistic Pony on Clip Skip 2 (First Setting in the "Sampler Settings"-Group top right, check if it correctly generates an image for the pony model and then change the model to lustify and try to generate one too. If the lustify image doesn't look right, set the Clip Skip to 1 and generate again to check if lustify is trained on Clip Skip 1.
Currently the workflow does not support setting a different clip skip for the base model and the refiner BUT if the lustify model indeed needs a different clip skip setting you can click on the "Model Backend" node at the bottom in the Backend Area and then open the subgraph by clicking the button on the top right of the node. It'll lead you to a subgraph that handles the Clip Settings and there you can find the Clip settings for the Refiner.
https://i.gyazo.com/199c9719175e3e8ed87a01eacd7a0a21.png
You can then remove the connection to the value of the node and change it to -1 or -2 depending on if it's supposed to have clip skip 1 or 2.
Hope that helps :)
I really like your workflow but Im keep failing with those changes, check out this one https://civitai.com/articles/6651/refining-ponyillustrious-models-with-sdxl15-for-full-realism-the-right-way maybe next time you can consider refining by sdxl, it seems to be popular way to add realism to pony, thank you for your help.
@sssebinho303 thanks for the buzz!
Can you let me know what you're struggling with or whats not working?
It should be possible to use it somewhat like described in the guide you've sent me with the options of the workflow.
@vslinx Im not really sure if it's the clip skip that causes the issue (chat gbt says is 1 for sdxl and 2 for pony), and also im not sure which node to detach and how to set -1 for refiner. I was trying to add "int constant node" I various places but every time when I enable refiner it fails (shows me that funny drawing from "load style image" as result).
@sssebinho303 i don't quite understand, the "load style image" is just for IPAdapter which you won't need for simple image generation.
Are you using the TXT2IMG or IMG2IMG workflow? I think you're having a general issue with the understanding of the workflow right now.
I have downloaded both checkpoints you mentioned and they're actually working with with clip skip 2 for both, so you don't have to do anything. I wouldn't really listen to what chatgpt says, it doesn't really have knowledge apart from the base architecture it's built on.
I have just thrown in your two checkpoints - lustify as the refiner model and cyber realistic pony as the main model, kept all settings as default and hit generate with this prompt: "egirl, 1girl, grey hair, underwear, looking at viewer, lying on side, black thighhighs, latex, thong, black gloves, ass, black tube top, sexy, amateur photo, dim lighting, plush toys, curvy, (dim lighting, purple lighting:1.2), fairy lights,"
and got these results:
https://i.gyazo.com/45918f3534e7b039dc4dd01af164724f.jpg
You don't have to add anything or change anything in the workflow, by default the Clip Skip (see Sampler Settings) is already set to two which works fine. The refiner does 20 steps at base meaning it'll do 10 steps on the lustify refiner checkpoint when you're doing 30 steps.
If you can show me screenshots or the console error that's showing i might be able to help you but with the explanations you've provided so far i don't even know what your problem is :/
@vslinx if it works for you then problem is not the clip skip. there is a hope. (Chat GBT fooled me many times)
so only Cyberrealistic pony: https://drive.google.com/file/d/1E0f4wAjIkC0EgUueJL7o0CAU5H4578Cq/view?usp=sharing
then when I turn refiner on (lustify): https://drive.google.com/file/d/1qYWQTPZLdDkYv1Xiaoj2xG3GEcvdfCcc/view?usp=sharing
And another problem Im getting is Adetailer :/ (when I turn detailer on):
https://drive.google.com/file/d/1skcN1aQsLKvbMs0TtdfJwyJvxUyacVcE/view?usp=sharing
Models are downloaded and showing on the list:
https://drive.google.com/file/d/12CpXQv9PjVjRL4j20u69ImhOFEoVrKsd/view?usp=sharing
hey @sssebinho303,
So cyberrealistic Pony successfully generates an image.
It looks like the error only comes from when you're using refiner.
The issue with the refiner stems from it receiving the seed where it's supposed to have no input at all since the "add noise" value is a boolean.
I don't know if you've tried to change anything or what error you've given chatgpt but this is definitely not an issue with clip skip. In the future always provide the error log as it's really the only way to figure out whats happening.
The second problem comes from a version incompatibility. Your comfyui is outdated as well as some of the nodes probably. Make sure everything is up to date, ComfyUI has recently changed the method name from .execute to .apply which is why you're seeing this error.
Your node seems to be on a newer version which is why it's trying to call the apply method while your comfyui is outdated and is still expecting things to be called with the execute method.
I'm not super familiar with runpods but you will have to update your comfyui installation.
Usually you do that by going into the terminal where your comfy is installed and executing "git pull" followed by simply starting your comfyui. Maybe your runpod has a built-in way to simply update comfy. Make sure you also update the frontend component of your comfy installation on runpod.
I assume that the KSamplerAdvanced issue you're getting with the refiner is also because your comfyui did some mistakes when loading the workflow. It's very likely it has connected the seed to the wrong pointer on the node itself. So update your comfy + all your nodes and then re-import the workflow from the file downloaded here. If you're still running into the sampler issue you can let me know and i'll show you how to fix it if necessary :)
@vslinx I've updated everything, adetailer works now. I downloaded workflow again, didn't change anything, literally run your prompt. Cyberrealistic solo was ok but with refiner, still the same error: (from logs) got prompt
Failed to validate prompt for output 274:* KSamplerAdvanced 546:576: - Value not in list: add_noise: '24334709765024' not in ['enable', 'disable']* KSamplerAdvanced 546:575: - Value not in list: add_noise: '24334709765024' not in ['enable', 'disable']Output will be ignoredFailed to validate prompt for output 532:@sssebinho303 Okay i have recorded a little video for you.
https://drive.google.com/file/d/1DznmLxm9nWTq_kX_tMLBXvE2GMkx1LvD/view?usp=sharing
here you can see where you find the technical stuff behind the refiner, from your error it looks like for some unknown reason your comfy is connecting (most likely) the seed variable to the "add_noise" value of the sampler.
If it is, make sure to grab the little dot and move it up to the seed instead. Just make sure nothing is connected to the "add_noise" true/false switch.
@vslinx unfortunately its still the same. here you can see the nods, Ive moved them so its easier to see what's connected. https://drive.google.com/file/d/1lgdp6frK4yZmajM7dkyZgZGoDeqo6DR4/view?usp=sharing
logos are still: got prompt
Failed to validate prompt for output 227:
* KSamplerAdvanced 546:575:
- Value not in list: add_noise: '24334709765024' not in ['enable', 'disable']
Output will be ignored and after that ignoring everything, the whole generation last less than 2 sec, not even loading anything from the main model
@sssebinho303 hmmm check the second sampler, the "Refiner Sampler"-Node if it's the same.
edit: just checked and yes the 575 id should be the refiner sampler (below the sampler)
@vslinx You are a Legend! its good now. it wouldn't work with any refiner because when I load free workflow, somehow instead of of enable/disable those nodes have some random number there! now when Ive changed to enable it works 🙏
@sssebinho303 nice i'm glad it works now!
I'm still confused why your comfyui decided to connect the nodes differently when loading the workflow, but i guess that's some mysterious comfyui bullshit :'D
@vslinx connection was good, just two hidden nodes were coming always with pre-set random numer instead of enable/disable. noise had default value 24334709765024. pure magic
Hello.. its great workflow! but sadly new update had problem about differental diffusion.. can you help me to fix it? so far as i know its added before facedetailer nodes.. but yours workflow its too complicated to try to change XD.
hey @saytolol12 sorry for the late reply, i didn't get a notification!
The problem with differential diffusion was fixed by Impact-Pack around 1 1/2 weeks ago.
So if you've updated Impact-Pack(+Impact-Subpack) to the newest version it's most likely that your comfyui is outdated. Make sure you update your comfy AND your custom nodes to the newest version! 🥰
@vslinx i see thanks for information.!! i'll try it
This is my 'daily driver' for Pony/Illustrious. Would love Wildcard support! (also i changed the lora loader to the lora manager one)
hey @slln,
happy to hear!
The workflow supports wildcards since v4.4! You can find how to use it here:
https://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/tutorial/ImpactWildcard.md
Just type your wildcards into the prompt fields according to the Syntax and it'll work 😎
i feel like i need to pay for stuff like this, amazing work!
everything's free over here @blackjuker haha 🥰
Thank you so much for the buzz you've send me! i'm glad you're enjoying the workflow ♥️
i have everything turned off, so just bare bone, but it generates 16 gray images.
batch size is 1, and disabling or enabling the vae does nothing.
hey @Taimanin001,
what workflow are you using - TXT2IMG or IMG2IMG?
What is your console saying?
Are you looking at the Before/After image preview nodes or at the pythongoss/comfy image queue(or image feed) ?
The image feed will always show every single image generated throughout the process of image generation, including each node that provides a preview. Since there are 8 previews for the detailers you'll automatically have at least 8+ images during a normal image generation.
You can change this behaviour by going into your comfy settings > pysssss > Turn "🐍 Image Feed Display 'SaveImage' Only" on. This way you'll only have images that have been saved in your image feed.
Let me know if you see an image in the "Preview after Upscale/HiRes Fix" and "Preview after Detailing".
@vslinx hi, i really don't know what was causing this problem since it shows no error in the console, but i did a fresh installation to comfyui this morning, and now it's working fine 😅.
Thanks for the reply
@Taimanin001 glad it's working now! and sorry you had to reinstall comfy :/
It sometimes can be a hassle!
Fantastic work. I've spent a year discovering the bits and pieces for a year, my head would have exploded if I had tried to use something like this before I knew what all the individual parts did. One note on the hrFix subgraph- as imported there was a mismatch on the sampler/scheduler links. I've had this happen elsewhere so I knew what was going on, but with the new system it might be really hard to debug for a tyro. I don't get a red outline on subgraphs when they halt.
Thanks a lot @Ponder_Stibbons ! ♥
I always try to make the workflow as easy as possible, but with the limitations of comfy it's not always possible to show everything in an easy and accessable way and it's even harder to make the technical parts still somewhat accessible for people who like to tinker :/
I always try to label everything incl. inside the subgraphs - don't know if that really helps anyone though haha
Yeah i'm not happy either with the error displays from the new comfy versions, i think it's big design flaw not to flag the subgraph the error happened in, at least groups did that before they got deprecated. Previously you were at least able to use the node id from the error log to search for it in the NodesMap and get to the failed node easily by a double click even if the node was in a subgraph but now they're only showing nodes that are in the current graph you're in. It's like every update they make debugging less and less accessable 🥲
Sadly i've also experienced it a few times that for some reason comfy upon loading a workflow starts mismatching subgraph inputs - no idea why that happens as it's not reproducable but i hope they fix that in the future!
I'm new to this whole thing but in the last 2 weeks I've tried a hundred different workflows on here. This one just makes sense and just works. Thank you.
ayyy @TootyPenguin i'm happy to hear that!
Thanks for the feedback 🥰
I can´t get it to work. Every time i try to generate the terminal throws the attached problem. Why are those inputs open when i load the workflow.
Failed to validate prompt for output 226:
* CLIPTextEncode 679:529:
- Required input is missing: clip
* ACN_AdvancedControlNetApply_v2 679:520:
- Required input is missing: negative
* KSampler 679:521:
- Required input is missing: model
* VAEDecode 679:524:
- Required input is missing: vae
* VAEEncode 677:650:
- Required input is missing: vae
* KSampler 677:651:
- Required input is missing: model
* VAEDecode 677:653:
- Required input is missing: vae
Output will be ignored
Failed to validate prompt for output 225:
Output will be ignored
Failed to validate prompt for output 29:
Output will be ignored
Failed to validate prompt for output 224:
Output will be ignored
Failed to validate prompt for output 667:
Output will be ignored
Failed to validate prompt for output 231:
Output will be ignored
Failed to validate prompt for output 230:
Output will be ignored
Failed to validate prompt for output 229:
Output will be ignored
Failed to validate prompt for output 274:
Output will be ignored
Failed to validate prompt for output 227:
Output will be ignored
Failed to validate prompt for output 228:
Output will be ignored
Failed to validate prompt for output 457:
Output will be ignored
hey @cyrax2212568,
it looks like you're missing the cg-use-everywhere node or it might be outdated.
make sure your comfy + nodes are up to date
https://github.com/chrisgoringe/cg-use-everywhere
@vslinx i already checked that. cg-use-everywhere is installed in the latest version, comfy,too. Anywax u tryied to update again, but no difference. Still the nodes seems not connect over "everywhere"
@cyrax2212568 cg-use-everywhere did not work for me either. Decided not to use it
@cyrax2212568 and @2P2 i updated my comfy and the nodes now and i can't replicate this problem. can you check the "Checkpoint Everywhere"-Node in the "Sampling"-Group at the bottom and tell me if the stuff there is connected?
https://i.gyazo.com/73b43caa279ad0b5fb9608e3840582de.png
I had an issue with cg-use-everywhere in the past and i'm currently debating removing it anyways. In the past it was convenient since people could easily add another lora or detailer but by now the workflow has gotten so complex that it might not be worth keeping this as a feature.
I actually also just got this issue when trying to generate something 😓it's a shame bc the workflow works beautifully otherwise. This issue seems new (+ something else where hires fix and the face detailers are acting as though they aren't enabled when they are). It makes me wonder if the update to the new frontend did something? Idk I'm not super familiar with this stuff
@pinkmoth068 hm that is weird :/
I have the latest frontend checked out, all nodes and comfy up to date and it works fine :(
Maybe comfy shipped some shit again?
@vslinx "Checkpoint Everywhere" is connected here, but its like it doesn´t matter if they are or not. Even tried a complete fresh install of comfy in the latest version and updated all nodes, still the same :-(
@cyrax2212568 man that sucks… it does work for me :/ i‘ll download the workflow really quick and fix this issue and re-upload it
@cyrax2212568 @qek @pinkmoth068 can you guys test this now?
i have removed all use everywhere nodes including prompts and models:
https://www.mediafire.com/file/quyfbzzvr40ixv9/TXT2IMG-ADetailer-v4.5-vslinx.json/file
If this works for you i'll update the file :)
The following days i'll release 5.0 with manual inpainting and will remove it from there too.
@vslinx Yeah I'll test it as soon as I can, thank you so much!
@vslinx thank you very much for your fast help. The uploaded version works for me. Looking forward for the updated img2img version soon :-)
@cyrax2212568 i have updated both the img2img as well as txt2img workflows here on the page now. You can redownload them and they should work :)
The 5.0 for TXT2IMG will come at some point this week and IMG2IMG will come sunday/next week :)
Hope everything works for y'all now!
@vslinx Just tested it, it works for me too! thank you 😊😊
This is a wonderful workflow and it's helped me understand and use comfyui as a new user. However I'm having a lot of trouble understanding how detailer works and I'm wondering if I can get some help.
Face detailer specifically is changing faces drastically even at ludicrously low denoise strength like 0.05. I'm using the same facial tags as I use in my original prompt but it changes the face so much that it looks almost unrecognizable compared to the original Hi-Res Fix version.
From someone who is used to inpainting in Forge and able to constantly iterate until I get a result I like, this is confusing new territory to me and I would appreciate any help I can get.
Hey @Zisa,
i 100% get you - the detailers can be quite confusing tbh and are not as greatly documented as i often wish...
That's why i wrote down in the bottom of the workflow a little guide of what each variable in the detailers itself do, you could try playing around with that.
For starters i'd recommend checking and making sure that the denoise slider is actually connected to the detailer node. Sometimes some people have weird faulty comfy installations that remove some connections when loading a workflow.
The second thing i'd recommend trying is to keep the denoise very low like you do and leaving the prompt empty/deactivating the prompt. If that works fine and looks good, add a prompt but not too many tags.
In forge and webui you usually have the full image size as the context window, in comfy and more specifically the face detailer you set the context window yourself. I have written down a little bit about it in the bottom of the workflow right next to the detailer.
Try to keep the prompts very small, if you just want to change the eye color for example only prompt the eye color like "red eyes", if you prompt anything that is not part of the image, like the character name for example and that does not make sense in the set context window in the settings of the detailer, it'll try to create that prompt into the image even if it means creating an abomination.
@vslinx Thanks for the amazingly quick reply. I'll give that a try. When inpainting in Forge, I would keep relevant loras when doing passes. Would you recommend using style loras used in the original generation or only using a detail lora or loras designed for specific body parts?
@Zisa no problem 🥰
If the loras are just style lora's i'd keep them.
You can add body-specific loras if you want in the detailer lora area, otherwise the general loras are being used throughout the whole process including the detailers. I'd refrain from using character-specific or body specific loras in the general lora loader to prevent detailers from getting these loras too :)
So just using style loras should be fine!
@vslinx It's actually very good to know that the general loras are being used at every step of the generation process. I'll take your advice and keep playing with it. I'm still not getting great results but removing the face detailer prompt at least kept the same face structure, it just removed all the makeup. So I just need to figure out how to fine tune that!
@Zisa If that's the case i'd try to re-prompt it by adding one after another like "eyeshadow" or "makeup" etc.
If you want to make sure that your checkpoints actually understand what ur trying to prompt, make sure to use the danbooru tags they're trained on.
You can do so by activating autocomplete with a danbooru tag list, you can find how to do it here in my guide:
https://civitai.com/articles/15480#autocompleter
Some people don't have file permissions to do load the file, if that happens you can instead press Ctrl+A to mark all contents of the file and then copy and paste it into the window in comfy instead.
@vslinx I'll definitely take a look at the article and getting the autocompleter. I tend to just double check danbooru often as a habit but I know a lot of the prompts by heart at this point.
I guess when it comes to face detailer my confusion remains on how exactly the detailer prompt interacts with the image. You suggesting to leave the prompt blank was actually a huge help. Now I have a building block to use to slowly add things until I get a good result. For most of my generations, I'm using the same tags to make a consistent character, so if I get a good detailer set up, I won't have to change it too much going forward.
@Zisa Okay, i'm glad i was able to help a little!
The 3 important factors really are the models you use and if it's bbox or segm, for more accurate changes with less influence from outside i'd say segm is better than bbox since it's only affecting the actual area. Next would be the prompt, keep it as light as possible and together with a low denoise you should be good as long as you use the same loras and checkpoint. Last would be the context window in the FaceDetailer that basically tells the detailer inpainting job to look around the area of the mask in an X factor to understand what the context of the full(er) image is. And obviously you'll have to try around a bit with seeds.
If you need anything more let me know (✿◡‿◡)
Good luck! Sometime this weekend/start of the next week i'll release the 5.0 with manual inpainting, so you can also start using that if the automatic detailers don't work to your full satisfaction 🥰
@vslinx I think I have something to work with here thanks to your help. I'll definitely be trying out 5.0 when you release it though. Your workflow has helped me switch to comfyui and learn how it works without being completely overwhelmed. Thanks so much for your help!
@vslinx I guess my final question is about the eye detailer. I'm getting good results with the denoise setting and prompt I'm using, but currently it only ever seems to be detecting one eye. I'm using Eyes Detection (Adetailer), the first one you recommend in the workflow.
I tried it both under bbox and segm. I tried lowering the detection threshold down, even to 0.3, but it still doesn't seem to want to detect both eyes. If I can solve this problem, I think I'm really close to having a set up that works for me.
@Zisa yeah the detection is completely based on the ultralytics model you use.
Eyes can be specifically tricky sometimes especially when talking about side eyes or forms of irises the models werent trained on.
If that model is not detecting the eyes i'd try switching to a different one and see if that manages to do it.
Here are some that i use regularly:
https://civitai.com/models/178518?modelVersionId=582143 (you can try the paired and it'll basically recognize the eyes including the bar between them or the single ones. The paired version usually results in pupils etc. having the same size which is preferable)
https://civitai.com/models/334668/eye-detailersegmentation-adetailer
https://huggingface.co/camenduru/ultralytics/blob/main/PitEyeDetailer-v2-seg.pt
@vslinx Yeah I swapped to the Eyeful | Robust eye detection one and that picked up both eyes without an issue. Maybe I'm too dumb to figure out the individual one but it seems to require 2 mask layers that get merged together and I can't seem to figure out how to do that with comfyui.
But again, thank you so much for the help, I'm definitely on the right track and you kind of cleared a mental block for me and I've fixed a lot of my issues by resetting my process. Lifesaver!
@Zisa the single eye detection should still work in one swoop in the default workflow's implementation. It simply generates two singular masks and inpaints them individually. So it's less good for consistent eyes but very useful if you really only want single eyes instead of a mask.
Very happy to hear! wishing you luck on your endeavours! 🥰
Sorry if silly question but where do you find all those tiny detailer files? Just rewrite the names and google them?
For instance the adetailer for nose doesn't come up.
Heya @Arararagi,
i mostly find them here on civitai by filtering or on huggingface by recommendations.
You can search for them by setting the filter on models to "Detection" like this.
The nose detailer was here on civitai too, but the owner removed it. Which is why i re-uploaded it and linked it in the workflow and here in the guide too.
It mostly works for anthro noses but is not that good for human noses.
It's still not working, even if I use the workflow you updated several days ago. Any idea?
Failed to validate prompt for output 227:
* KSampler 546:82:
- Return type mismatch between linked nodes: scheduler, received_type(['simple', 'sgm_uniform', 'karras', 'exponential', 'ddim_uniform', 'beta', 'normal', 'linear_quadratic', 'kl_optimal', 'AYS SDXL', 'AYS SD1', 'AYS SVD', 'GITS[coeff=1.2]']) mismatch input_type(['simple', 'sgm_uniform', 'karras', 'exponential', 'ddim_uniform', 'beta', 'normal', 'linear_quadratic', 'kl_optimal', 'bong_tangent', 'beta57'])
With Detailer on,
* FaceDetailer 19:
- Return type mismatch between linked nodes: scheduler, received_type(['simple', 'sgm_uniform', 'karras', 'exponential', 'ddim_uniform', 'beta', 'normal', 'linear_quadratic', 'kl_optimal', 'AYS SDXL', 'AYS SD1', 'AYS SVD', 'GITS[coeff=1.2]', 'LTXV[default]', 'OSS FLUX', 'OSS Wan', 'OSS Chroma']) mismatch input_type(['simple', 'sgm_uniform', 'karras', 'exponential', 'ddim_uniform', 'beta', 'normal', 'linear_quadratic', 'kl_optimal', 'bong_tangent', 'AYS SDXL', 'AYS SD1', 'AYS SVD', 'GITS[coeff=1.2]', 'LTXV[default]', 'OSS FLUX', 'OSS Wan', 'OSS Chroma']).
Hey @harlemrayx,
please check out the error in the FAQ of the guide.
Hope this helps 🥰
@vslinx Get solved. Thanks a lot
sorry for a really stupid question, but how do i get it to save the images to the output folder and not the temp one and to save the metadata linked to the images? ive been using stability matrix as a workflow for a while and only recently decided to switch to comfyui, so im struggling really hard (｡╯︵╰｡)
heya @chloeChu,
there are no stupid questions - all good :D
The images should already save to the output folder and have metadata of the generation included. If the images are not saved to the output folder it'd mean that the workflow didn't run all the way through and you should see a error in your console that tells you what happened.
If you show me that error i could help you figure out what the issue is (✿◡‿◡)
@vslinx i reinstalled python from scratch and that seemed to help, thank you, this workflow is actually the best one i could find (´ ω `♡)
@chloeChu very happy to hear!
Glad you're enjoying the workflow ^-^
This has been an awesome workflow so far, but for some reason my Image Saver box is highlighted red around it and I don't see the final results showing up in my output even though I see it in the preview after detailing box. Any ideas? Do I need to manually configure anything in that box?
Hey @zhalor,
glad you like the workflow!
Usually there shouldn't be any issues with the image saving, can you tell me what error you see in the console?
@vslinx Thanks for getting back to me! I am not too worried about it since I can save it via right click and save in the output frame. I don't see any errors in the console that really lead me anywhere. I will try another run later today to really scrutinize the console to see what I can find.
I do however have a question on how to get less washed out results. I am using the hi-rez fix and hi-rez color fix toggles, but the output tends to come out a lot more washed out than the original. I am using a v-pred model so I have that enabled, and I don't have an additional vae option toggled since I just want it to use the same vae as my original checkpoint. Any ideas? Or is it just the nature of hi rez fix? I don't have the same problem when upscaling using Ultimate SD Upscale.
@vslinx This is the error I am getting regarding the image saver
Failed to validate prompt for output 457:
* (prompt):
- Exception when validating node: tuple index out of range
* Image Saver 457:
- Exception when validating node: tuple index out of range
Output will be ignored
@vslinx Sorry for the bombardment, but it seems to be tied to the img2img transfer portion. I didn't have that enabled, so it was just lighting up, although even with it toggled on it still occurs, so I am not sure.
@zhalor It sounds like the hires color fix is not really being applied either if the image is washed out. Since that's what it's there for.. :/
These tuple index out of range errors have appeared when subgraphs were introduced, could it be that your comfyui and/or plugins are outdated?
It should only appear if there's an incompatibility.
Can you check your Comfy version and frontend version? When you start comfy there should be a few lines that look like this:
"Python version: 3.12.3 (tags/v3.12.3:f6650f9, Apr 9 2024, 14:05:25) [MSC v.1938 64 bit (AMD64)]
ComfyUI version: 0.3.68
Initializing frontend: Comfy-Org/[email protected], requesting version details from GitHub...
[Prompt Server] web root: G:\ComfyUI\web_custom_versions\Comfy-Org_ComfyUI_frontend\1.28.8
[Crystools INFO] Crystools version: 1.27.4
[Crystools INFO] Platform release: 11
[Crystools INFO] JETSON: Not detected.
[Crystools INFO] CPU: 13th Gen Intel(R) Core(TM) i9-13900K - Arch: AMD64 - OS: Windows 11
[Crystools INFO] pynvml (NVIDIA) initialized.
[Crystools INFO] GPU/s:
[Crystools INFO] 0) NVIDIA GeForce RTX 4090
[Crystools INFO] NVIDIA Driver: 580.97
[ComfyUI-Easy-Use] server: v1.3.4 Loaded
[ComfyUI-Easy-Use] web root: G:\ComfyUI\custom_nodes\ComfyUI-Easy-Use\web_version/v2 Loaded
ComfyUI-GGUF: Allowing full torch compile
### Loading: ComfyUI-Impact-Pack (V8.25.1)"
@vslinx
Here is what I get.
Python version: 3.13.6 (tags/v3.13.6:4e66535, Aug 6 2025, 14:36:00) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.3.68
ComfyUI frontend version: 1.28.8
[Prompt Server] web root: F:\A.I\A1111\ComfyUI\python_embeded\Lib\site-packages\comfyui_frontend_package\static
[Crystools INFO] Crystools version: 1.27.4
[Crystools INFO] Platform release: 11
[Crystools INFO] JETSON: Not detected.
[Crystools INFO] CPU: 13th Gen Intel(R) Core(TM) i9-13900K - Arch: AMD64 - OS: Windows 11
[Crystools INFO] pynvml (NVIDIA) initialized.
[Crystools INFO] GPU/s:
[Crystools INFO] 0) NVIDIA GeForce RTX 4090
[Crystools INFO] NVIDIA Driver: 581.57
[ComfyUI-Easy-Use] server: v1.3.4 Loaded
[ComfyUI-Easy-Use] web root: F:\A.I\A1111\ComfyUI\ComfyUI\custom_nodes\ComfyUI-Easy-Use-main\web_version/v2 Loaded
as for the colorfix, when I run it off it is substantially more washed out, but when it is enabled it still loses the vivid colors from the original image generation, so maybe that is just an expectation of hi-res fix?
Appreciate all of your help! :)
@vslinx I stand corrected. I adjusted the sampler I was using and it seems to be fixed!!! Thanks a ton for helping me through this.
@zhalor the hires fix will always wash out the original a little bit, even more so if the VAE doesn't match the original, BUT the color fix should always fix it.
If the problem still persists you can go to the bottom of the workflow where you can find the "Color Fix Sub", you can enter it by clicking the little icon in the top right corner like here.
In there you will find this node where you'll be able to increase the strength of the color fix.
This will apply the color correction more aggressively.
Has changing the sampler also fixed the issues with the tuple?
If not you should check your custom nodes manager and see if there are any updates for any of the nodes. It might be that one of the nodes still pass a tuple since it hasn't been updated since comfy's latest rework.
@vslinx Great to know about the color fix! I did a re-install of Comfyui with fresh nodes and that seemed to go away, so we are in good shape now. Again, I can't thank you enough. You are awesome.
WHY the sampler settings are all blank,such as size,batch size,steps？
It`s ok,mxtoolkit is conflict with mixlab.https://github.com/Smirnov75/ComfyUI-mxToolkit/issues/28#issuecomment-2603091317
Wow this sucks @Xyxxxb !
Seems like it was reported in January already and the developer of mixlab still hasn‘t fixed this. Thank you for letting me know! ❤️ This way i can help others running into this issue
Hey i just got this workflow and it is MASSIVE! Just as i started looking around it to learn i realized something must be wrong because 80% of my workflow has Red highlighting around every node. Important to note that i've been using ComfyUI for a few days now and i actually have a private working flow with over 10 loras and everything works great (though with a few issues im still learning). So what is all the weird Red highlight boxes around every node? Its literally around every single node, 80% of the workflow is marked red and some of the lora nodes etc have Text boxes so big its going into other nodes.
Edit: Nevermind i figured out i gotta install a bunch of custom nodes.
heya @Arookb,
yeah like you mentioned you sadly have to install a few custom nodes :/
You can find all of them in the description, or install them easily through the comfy node manager.
Comfy in it's base form sadly doesn't allow for most of these features i have implemented.
But whenever something new becomes available in the core comfy, i'll always replace the functionalitys with core functions to make it more lightweight.
I hope you'll be fine with installing the nodes and that you'll come to enjoy the workflow ^-^
Hi, great job first of all, I enjoyed using this workflow very much. My question is, how do I prompt with multiple characters? I want two different characters in certain positions but how do I make it work?
Heya @fortniayti,
Sadly with SDXL Based models like Pony, Illustrious, NoobAI and SDXL itself it's mostly a gamble on how good you can get two or more different characters in specific positions.
Since they're not trained specifically with an understanding of this concept in mind.
The easiest way to do that is to just generate two people in whatever form you're intending to and then using inpainting to draw a mask over the characters to change the character it's displaying. But that obviously requires some manual steps along the way.
I'm working on the 5.0 right now and i'm in the last steps of only adding the explanation nodes right now, so it'd be something you'd be able to do soon.
There are other workflows that are specifically made for that purpose alone like this here.
Obviously you can also just play the lottery and hope that the Checkpoint/LoRA's you're using know well enough which character has which features. So you could try a prompt like this:
<quality tags>, <character tags>, <action>,
So for example you could do:
"masterpiece, best quality, absurdres, amazing quality, " (start quality prompt)
"inoue orihime, orange hair, brown eyes, long hair, large breasts, straight hair, blue jeans, turtleneck, white sweater,
kuchiki rukia, black hair, purple eyes, short hair, hair between eyes, small breasts, black jeans, turtleneck, black sweater,
2girls, hug, smile, " (Positive)
Which would generate this image for example.
Obviously it's a bit of a trial and error, but if you describe the characters and the action well, it'll usually work out quite well too. That being said it's heavily dependent on how well the checkpoint and/or the LoRA you're using know the character and/or the position you're trying to put them in.
Hope this helps 🙏
@vslinx can't wait for 5.0!
Thank you so much @blackjuker for your continous buzz donations! ♥️
Since i'm only in the process of writing the final note you could also download the current Work-In-Progress version here on my huggingface.
I should have the time to write the remaining tutorial tomorrow and release it tomorrow evening 🥰
@vslinx That's Lovely! thank you! take some more buzz~~~~ <3
@blackjuker you're buzzing man! 😂♥️ thank you a ton!
Just a heads up in the 5.0 version IMG2IMG, the following error occurs : Failed to validate prompt for output 225:
* CLIPTextEncode 679:529:
- Required input is missing: clip
* ACN_AdvancedControlNetApply_v2 679:520:
- Required input is missing: negative
* KSampler 679:521:
- Required input is missing: model
* KSamplerAdvanced 679:692:
- Required input is missing: model
* VAEDecode 679:524:
- Required input is missing: vae
* VAEEncode 677:650:
- Required input is missing: vae
* KSampler 677:651:
- Required input is missing: model
* VAEDecode 677:653:
- Required input is missing: vae
my noodle brain can't figure out why
heya @blackjuker,
the IMG2IMG version of the 5.0 is not ready yet - i wanted to finalize the TXT2IMG workflow completely before porting the changes to the img2img version since that always requires a few different changes especially now with the introduction of the inpainting.
That implementation has changed A LOT of the original way the workflow worked to allow for a smooth branching of these processes.
I had not yet pushed the newest working changes for the 5.0 where i had removed the anything anywhere nodes since they became obsolete with the overhaul of the technical parts i did for the workflow.
I have just pushed the changes for the 5.0 now, so you could download it now :D
But don't be surprised that the inpainting is not yet implemented in there 🫣
This workflow is great!! however After disabling IMG2IMG transfer toggle the workflow broke and it won't process the pictures in normal mode anymore, seems like the checkpoints aren't even running and just outputs the same input image,, and for some reason even when using a fresh workflow now shows IMG2IMG transfer enabled by default when that was not the case originally. seems like the toggle is stuck and when disabled it doesn't enable back normal IMG2IMG mode. please help.
Heya @HazelYK,
please read the note above the IMG2IMG Transfer Group.
The 4.5 version comes with img2img transfer activated automatically, so that is correct ^^
The reason why it doesn't work now is most likely due to the technical limitation mentioned in the little note above the IMG2IMG Transfer Group.
I have by now implemented a fix for this behaviour that'll be included in the v5.0, if you want to you can already download it here.
The remaining part until it's release is for me to implement the inpainting i have already finished for the txt2img workflow.
You're the best thank you!!
i‘m glad if that helped you @HazelYK! ❤️
Thanks a lot for the buzz donation! 🫶🏽