ComfyUI Detailer/ADetailer Workflow

Hey tiru92,
thanks for the heads-up and feeback! Glad you're enjoying the workflow and guide ♥

Have been using the mouth model for 5 months now and seemingly the creator added it last month into the title + answered it in a comment ^^

tiru92 no need to thank me! :D

Usually people only use one or the other, i haven't really seen anybody use multiple UNLESS you do manual inpainting of specific bodyparts like the hand - then some people do a second run-through of only the hand with an openpose controlnet of a hand in the pose that they want the hand to be for example. But yeah the scene moves really fast, so i wouldn't be surprised if using multiple might've been common practice in the past.

I think it can also heavily depend on the model you're using. I know that SDXL Models didn't have ControlNet Models for a while and therefore chaining two or more ControlNets with different processors like Canny, OpenPose, LineArt etc. in a chain might've gained better results than doing any one of them alone since they weren't trained well on them.
But nowadays you can easily use just one of them and adjust the strength accordingly, chaining multiple will most likely fry your results as they're being parsed into the CLIP that is responsible for the understanding of positioning in your generation.

tiru92 my pleasure! :)
Usually canny always give me the results i need, like you said it's just a matter of tweaking strengths.

In-Painting is the next big feature on the roadmap.
You can always check my intended next upcoming features in the description - they're always listed at the top. 🫡♥️

tiru92 i'd have to check how to connect them :)

Sadly i can't access your article, you could also post an image that you generated with that workflow, by default the workflow will be saved with the image. So if you have uploaded it i can just drag and drop into my comfy to check it out :D

(i also already developed a prototype but some other things have gotten in between like subgraphs and some restructuring)

tiru92 your inpaint area looks great already! if it's okay with you i'll probably steal a bit from it for the v4.5 which will have inpainting included then! 🫡♥️

tiru92 yeah i figured that'd be the case 😂

That's why it's at the end of my to-do list haha

hey @tiru92,

you are correct in your thinking - the detailing is the last part of the workflow and happens after upscaling & the hires fix.

There are different approaches with detailing where you either do detailing before the upscaling/hires fix, where you do it after (like here) or where you do one with bigger denoise BEFORE upscaling to fix anatomic issues and then another one after everything is done with a low denoise to touch up on wanted areas.

When i made this workflow i took into consideration that nowadays most pony/illustrious/SDXL checkpoints will not need major anatomy fixes. Another reason is that upscaling will inevitably introduce artifacts, which i also want to eliminate with detailing. Last reason was also the introduction of hires fix, which will re-sample the image anyways and mostly fix major anatomy issues (depending on denoise value ofc). So if detailing is done BEFORE the upscaling/hires fix it'll mostly just be to fix major anatomy issues instead of actually upscaling/improving the body part for the final result.

One way to achieve what you want to do could be to not use upscaling or hiresfix in the txt2img workflow and do the detailing there, then use that picture in the img2img workflow with upscaling and/or hiresfix to finalize. I know that's not the ideal approach to switch workflows but i think for now that's the way to achieve that.

You could add another upscale node(and leave the current one bypassed) after the detailer and feed that into the hiresfix node instead. This way the detailing will happen first and then it'll run into the upscaler and then into the hiresfix if you'd prefer using the detailer to fix anatomic issues.

I will see what i can do to maybe optionally move the detailing pre-upscale.

@tiru92 yeah i gotchu, some of the older illustrious checkpoints still had some issues :/ mostly because they were overtrained. Sometimes i used SDXL loras like the detail tweaker to fix it since it doesn't have an inherited style bias: https://civitai.com/models/122359/detail-tweaker-xl

You could try to increase the denoise of the hands and explicitly prompt what you're trying to do with the hands, that usually really helps. If you're trying to do a peace sign for example you could prompt "v sign" into the hands prompt. I never recommend actually writing "hands" since in no dataset that is ever trained in any checkpoint will there ever be hands explicitly tagged, so what usually works really well for me is to prompt "fingernails" since it's often extra prompted if they're prominent. Through denoising with a higher value (0.38-0.5) you'll usually get good fixes this way.

If the major problem was a unmatching color you could try using a color fix again on the final result like i did down in the "Color Fix"-Group or increase the feather of the detailer, it might fix the issue with wrong colors.

@tiru92 that is sadly always completely dependent on the detection model you're using.
You can try lowering the sam_treshold/bbox_threshhold (depending on segm or bbox model) and see if that works. The threshold is responsible for the confidence cutoff where the detailer decides if that thing is still what it's trained on. Usually to leave it high to prevent false-positives, but you can try lowering and see if the second hand also gets detected.

Heya @tiru92,
no worries - questions never bother me :D

I have tried to replicate it but couldn't. I can clearly see in your screenshot that the ControlNet Pose Preview is indeed showing the loaded ControlNet Pose and is using that during ControlNet instead of a pre-processed version.
Have you activated the PreProcesser in the "IPAdapter / ControlNet"-Group? It's sadly too far left in your Screenshot so it's not in the visible area for me to confirm.

If you have activated it could you rightclick somewhere in an empty space select "Workflowm Image" at the very bottom and then "Export" > "png" and upload that image to mediafire or something? That generates a image of the whole workflow with the actual workflow attached so i can check what's wrong.

@tiru92 thanks for providing all the data!

After checking your workflow i noticed that even though you have Pre-Processing activated, all nodes inside the Subgraph are bypassed. I don't know if you've done this manually as you've also bypassed the Refiner (not necessary, unless you switch from Normal Sampler to Refiner sampler the functionality is not used and the switch actually happens in the Sampler subgraph) OR your ComfyUI for some reason is ignoring the state of nodes inside when bypassing.

You should actually be able to simply fix this issue by simply de-activating and activating the "PreProcessing"-Function again. Also make sure the frontend and comfyui versions are up-to-date to make sure that everything works the way it's intended.

I also recommend not bypassing things like groups manually instead of using the switches because the workflow is built around the switches and their groups content having the exact same bypass/mute state. Manually un-/bypassing nodes inside groups or groups completely can mess with their state inside comfyui, that's why i've implemented switches for everything that is optional - since they handle the bypass state for all nodes inside the group INCLUDING inside all subgraphs. It could also be that pinning a subgraph, like you did, has prevented it's content from being un-/bypassed in a previous version of comfy/frontend, this should be fixed by now though since i couldn't reproduce that with the newest comfy&frontend version.

Here you can see the subgraph (clicking the top right icon). Here the nodes inside are bypassed even though the subgraph is not.

TL;DR:
Should be fixed if you simply press the "Enable PreProcessing" again to deactivate/bypass them and then press it again to activate them all.

@tiru92 no biggie!

if you manually bypass just the subgraph, does it's contents get un-/bypassed equally just like the subgraph itself?
I just did the same - downloaded the workflow from here, threw it in my own comfy instance and in my test instance that is newly generated every time i call it and in both cases using the bypass node as well as manually bypassing something it always also bypassed its content, as that is also the default behaviour (as of right now 0.3.56).

You could try by disabling the sampling subgraph, in there are 2 groups of which 1 is always bypassed (refiner), it'd be interesting to see if it bypasses it's contents or enables everything/disabled everything OR maybe preserves the initial state of one being bypassed etc.

It really does seem like your comfy seems to be bugged if manually bypassing and bypassing through the switches have the same behaviour of not changing the bypass status of it's nodes inside. I'd actually prefer if that'd be how subgraphs work since it'd be easier to write features around this behaviour :'D but right now that's not how they're supposed to work at least.

If the manual bypass behavior works and everything inside is also being bypassed it's very likely that your rgthree custom_node is either not updated or bugged, it'd be worth uninstalling the node and reinstalling it then. Otherwise if both the manual as well as the functional ones work the same it does seem like either your comfy has some issues OR another node you have installed is overwriting/interfering with the default behavior of this functionality :/

@tiru92 all good haha :D
Sometimes it's easy to miss the most obvious problems/solutions when you're looking at a lot of variables 🫡 Glad that it's working now and that it only was something as simple as an outdated node instead of a comfy bug! ♥

Hey slaad0,

sure no problem :)
Here is it for the TXT2IMG Workflow: https://www.mediafire.com/view/qge1j9w4kgqpjah/workflowTXT2IMG.png/file#
Here for the IMG2IMG Workflow: https://www.mediafire.com/view/abcgts768t6zumb/workflowIMG2IMG.png/file#

There's also preview images of the nodes more closed up here in the model page.
The next version is coming soon which will have most nodes cleaned up into the new comfyui subgraphs to hide most of the technical stuff that doesn't require user input and isn't customizable.

Hey cynseok,

the feed shows every single image created during a workflow. Meaning that every process during the workflow that generates any kind of image will show up in the feed - including the upscaling and detailers.

I can recommend disabling the feed completely in the settings as it‘s an old and flawed system that most don‘t use anymore since more complex workflows with more than just one sampler will always generate multiple images throughout the process.

Otherwise you should also be able to limit the images shown in the feed to only images that have been saved. You have to check your comfy settings for that - try searching „feed“ in the settings menu.

Hey dNerop,

thank you for the feedback and i'm glad you're enjoying the workflow! 🥰
As of right now a refiner is not part of the workflow. Mainly because i want to keep it as (somewhat) simple as possible and adding a secondary checkpoint means i'll have to also give options to inject lora's into that and have to do a lot of background calculation for the steps and add controls to determine how much of the process is being done in the first sampler and how much in the second etc.
Then there'd also be the issue of compatibility as some user's like to use SD1.5 or even Flux checkpoints for refining, which this workflow does not offer.

All in all i don't really feel like refiners do much since SDXL 1.0, a few users have done tests on how well refiners perform after the update from SDXL 0.9 to SDXL 1.0 and usually the results are either minimal or even negative. (see here for example)

Since Illustrious and Pony are based on SDXL1.0 i don't see too much of a benefit that would justify all the extra knobs and background modules to implement it :/ but i can recommend trying the hi-res fix. While it does not split the sampling process in two, it does increase the quality by denoising the previous input and if you want to, you can go down in the workflow into the HiRes Fix Backend and connect a different checkpoint if you'd prefer to do that with a different one than the original.

If you need some more help with comfy or with the workflow you can always ask :)
I also wrote a guide for the (IMG2IMG) workflow, but it largely also applies to the TXT2IMG in it's functions and usability. You can find it here: https://civitai.com/articles/15480/img2img-comfyui-all-in-one-workflow-guide

dNerop no need to thank me! :D

That might be something you'd have to try to accomplish with a LoRA then maybe if you want to use the workflow :/
Otherwise i also quickly whipped up a version for you that has a refiner, you can download it here: https://www.mediafire.com/file/u8nrxarpl49eqsm/TXT2IMG-ADetailer-v4.3%252Brefiner-vslinx.json/file

At the top you'll find a "Refiner" area, where u can select a secondary checkpoint + lora selector to add as many for the last few steps as you want.
The "Steps spent on the Base" will determine when the 2nd Sampler(Refiner) will take over. So make sure you have your Steps and the Steps spent on Base set correctly.

I wouldn't implement it like this into my workflows as of right now because i'd want to make sure that you can enable/disable the refiner and switch between the correct samplers accordingly, the version i made for you only works with the Refiner enabled, so you can't switch it off.

thank you dNerop ! 🥰
I'll try to be efficient (sometimes) 🫡
And always appreciate people sharing their work with the workflow! It looks amazing! Thanks a lot ♥️🙏

Hey sunioyu,

as oft right now there is no implemented dynamic prompt node to do the typical {red eyes| blue eyes|green eyes} prompting. I can see how conveniently it's implementable and check if that's something i'll add for the next version!

Currently if you have a VAE baked into your model it should be enough to connect the VAE of the model loader into the "Checkpoint everywhere"-Node in the Backend. I'll add a "Baked in"-boolean for the VAE for the next version - thank you for the feedback!

Heya PokieFish,
Thank you for the feedback! It was indeed a long process, but seeing the community enjoy it makes the effort worth it! 🫡🥰

I think i don't quite understand #1, do you mean to bypass the image selection process in the txt2img workflow? or to create multiple images in the IMG2IMG workflow?
If you want to generate batches of images in the TXT2IMG workflow you can increase your batch size, deactivate all the detailers and only leave the upscaling on. If you want to bypass the image selection before the processing starts you can go down into the "Backend - Model preperation"-Group and either Bypass the "Image filter"-Node or delete it completely. You can also open it, change timeout to "1" and then "on timeout" to "send all".
If you want to use the IMG2IMG Workflow only upscale you'll sadly have to wait for the v4.4 release, there you have a new node i developed to load multiple images at once instead of selecting solo images etc.

To disable the sound you have to go into the comfy settings and then to Image filter and deactivate "Play sound when activating". It's not in the guide because the guide is only for the IMG2IMG workflow and there the Image-Filter is not in use ^^

To #2 i can say that the v4.4 i'm working on right now will have dynamic prompts/wildcards implemented, including being able to choose from your preset-files. The file format/folder directorys will be from Impact-Pack, so this will be the logic of how it's implemented:
https://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/tutorial/ImpactWildcard.md

hey @MikeyArtis the positive prompt in TXT2IMG is simply "Positive"(+ start/end quality prompt if activated). In IMG2IMG if u have transfer activated it's the "IMG2IMG prompt" field.
If you're using IMG2IMG and have the transfer activated, the original will be checked and the prompts of the original will be copied into the image. You can always check your full prompt in the "Preview full prompt"-Node in the "Debug"-Group below the previews.

@MikeyArtis now i get what you mean! And i'm glad you're enjoying the workflow! ♥️

You can already add Lora's by text if you want, i sadly will not switch to the impact-pack nodes because they're overwhelming for new users. The next version (4.4) will support dynamic/wildcard prompts though so you'll be able to add files or wildcards to the prompts.

The reason why your image is not generating the way you want right now is because you're combining positive and negative prompt right now.
Meaning your full prompt also has your negative prompt in it. To fix this just connect the clip and "populated_text" output to the "prompt convert"-node instead. Then do the same with the negative to the "prompt neg convert". (You also have to change the model and clip order, but i've done this in the file below for you)

I have done this for you here:
https://drive.google.com/file/d/1p9WRzxZpcKy3kAiXJundIdzBdyhKIQrp/view?usp=drive_link

@MikeyArtis sorry, forgot to change file permissions - you should now be able to download it.
And no need to thank me :)

@MikeyArtis nice! I'm glad it's working! 🫡♥️

I know the switch from a1111 to comfy is hard and i hope this workflow will make it a tiny bit easier ^^

and thank you for the tip! ♥️

@Diemonde i have tried it with 16GB reserved, meaning only 8GB VRAM available.
It worked but obviously it was significantly slower. For the more VRAM heavy tasks(ControlNet & IPAdapter) i have implemented Low VRAM options to pre-downscale the input images for these.

I know of at least 1 other user that is actively using this workflow that has a 8GB VRAM Graphics Card but he has more RAM, so i'd say you should give it a try. If you're running out on memory on any of the tasks in the workflow you can let me know where that happens and i can give you some parameter changes that will decrease VRAM load. If you have triton installed you can also activate the Sage attention patch in the workflow which saves a bit of vram.

@Diemonde no need to thank me! and thank YOU for the kind words haha
i guess it's more of being an idiot with a lot of trial and error though

Hey @CitizenKing, does this happen on the txt2img or img2img workflow?

It could be that you‘re having the wrong image-saver custom_node installed. There is an old one with the same id that hasn‘t been updated in 2 years and therefore is deprecrated. Make sure you‘re using the right one from the description/guide/workflow-note.

If the issue still persists can you send me the log of your console?

Hey @fernanthecreator,

usually this only happens if you don‘t have all the nodes installed. Namely you‘re missing the ControlNet and IPAdapter nodes. Please check your console if none of them say (IMPORT FAILED).

If they do please check the instructions in the guide at the very bottom under the FAQ.

That should solve it.

Hey @warsound75,
seems like that's an issue with ComfyManager - i guess you have tried installing it through ComfyManager?

It's probably an issue with a version issue between your ComfyUI, your Frontend Version and/or your ComfyManager.

I recommend doing a "git pull" in your comfyui directory to update comfy, a "git pull" in your ComfyManager directory to update the manager and to add "--front-end-version Comfy-Org/ComfyUI_frontend@latest" as a parameter where you start comfyui (so probably "python main.py --front-end-version Comfy-Org/ComfyUI_frontend@latest").

If you don't want to update everything you most likely have to install Comfyroll manually without the manager. To do that simply remove the comfyroll folder in your custom_nodes folder, enter the custom_nodes folder through your cmd/terminal and do a "git pull https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes.git"

That should do it :)

@warsound75 it seems like your ComfyRoll might be installed but not working.

Does your console say "(IMPORT FAILED)" where ComfyRoll is being loaded?
Try opening your ComfyUI-Manager, go to Custom Nodes Manager and at the top left switch the filter to "Import failed", then if comfyroll pops up there click "Try fix", once it's done you can restart comfy and refresh the browser window where comfyui is loaded(even if you had closed it etc, reload regardless).

@warsound75 alright, sounds good! let me know if something‘s not working or you‘re missing something :)

hey @richardwolf99,

does it say "(IMPORT FAILED)" in the console when it loads rgthree?
If yes please follow the instructions in the guide under FAQ here: https://civitai.com/articles/15480/img2img-comfyui-all-in-one-workflow-guide

I just set up a new instance, installed all custom nodes and downloaded the workflow and it's working without issues for me. Are you running rgthree version "nightly 1.0.25082416"?

It could help to maybe delete it in the custom_nodes folder and to clone the rgthree node into it again and then going into the folder manually and to execute "pip install -r requirements.txt"

@richardwolf99 nice! sometimes Comfy-Manager can cause hiccups with bigger version jumps during installation/updates.

Let me know if you're missing something or have feedback :) enjoy!

hey @dlguddn119,

if there is no image in the output folder, it means that the detailing wasnt bypassed and failed - probably because of a missing model. Can you make sure you have the right model selected and check your console for any errors. You can also check the „Face“-Node in the Detailer Debug-Group and see if that preview image is completely black etc.

@arkhin1247 you'd be the first to experience that issue, have you selected models for the detailers and/or the upscaler? can you check your console that's running comfy for errors?