ComfyUI Detailer/ADetailer Workflow

Hey @elmejor9369,

this is the first time i have someone running into this issue :O What is your error log showing?

@BalmoAI if you're using google colab and install comfyui with the necessary nodes it should work just as if you're using it locally! 🫡♥️

@BalmoAI you‘re probably checking the image feed and not the actual output. In comfy workflows every „preview image“ node creates an image in the feed. You can turn this off in the settings of comfy to only show images from „save image“ nodes.

The previews are always done for every detailer as well as the previews for the pre hires fix and the controlnets, even if none of them are in use. This is so you can debug if anything isn‘t to your liking in the generation so you can see what to change.

The actual final image is the one that‘ll be saved in the output folder and which will also appear in the two big preview image nodes of the workflow. If there are no images there it‘d mean that the workflow didn‘t actually run and execute, if that is the case you‘ll have an error in your console, where comfy is running, that shows what went wrong. If you can provide this error i can help you if you don‘t understand what the issue could be :)

Hey @BalmoAI,

sadly i think there's not much that i can do to fix this :/
This is a problem exclusively with collab since your container get's thrown out once your time runs out and you have to set it up again.
The error you're seeing is because the file it's referencing is no longer in the input folder.
In ComfyUI the "select image" process works that way that you select images and these images upon selecting are being copied to your input folder inside comfyui.
Obviously if your image is deleted, removed or renamed the reference doesn't exist anymore. The preview might still show since the preview is loaded into the cache - but once you start the actual workflow the image isn't there so the error is being thrown.
In these cases dimensions are then also not being shown anymore since the information cannot be called on these files anymore.

It's also somewhat counter-intuitive but when you open a new colab/installation of ComfyUI, you always have to refresh the window where you have it open once by pressing F5 or clicking the refresh button.
Even across different url's and installations your browser persists all cached comfy data, so you might see a preview that's not there anymore and some nodes may not work because the shown nodes are still pointing towards your old colab/installation.

The workflow is already colab compatible and tested, the problem with colab is that you basically have to throw away the container after a few hours again once the free time runs out and then have to set it up again. Then most people don't clear the comfy cache and work on half cached versions that mismatch what's present on the hard drive of the colab :/

@BalmoAI no worries, hope it helps! ♥️

Hey @coxsirius766,
i'm glad the inpainting works well!
I'm mostly using checkpoints without a inbuilt VAE so i usually leave the option on since i mostly release it in the way the images i've created as examples were created ^^

So i hope that with the option to turn off the seperate VAE in the Model section it's understandable for most people :D
Thank you for your feedback!

@coxsirius766 do you mean generating multiple different images with multiple different resolutions in one go? I think i'm not quite understanding the request :D

85% of the workflow are hiding in subgraphs so it won't be too big, so it's gotten very complicated - therefore since it's such a niche feature i think it'd be easier for you to build it from scratch and make it do what you want rather than implementing it in this workflow ^^

aaaaah @coxsirius766 that makes sense :D

You can already do that with wildcards if you want. You can prepare prompts in a text file and seperate them by line, then load them in like __path/name__ into the promptfield and just increase the amount of tasks comfyui runs at once by increasing the number next to the play button.

Hey @clawsofdeath,
removing any links will result in you not being able to use some functions of the workflow.

By disconnecting the links to the sampling process you won't be able to use a global scheduler since the setting will not be passed down to other nodes anymore.
The reason you're getting this error can be found here and has to do with a custom_node you have installed that breaks the scheduler selection with any form of node that expects a scheduler list that is not the same as the one provided by the RES4LYF node.
Sadly the developer of this node stupidly overwrites the schedulers globally but doesn't do so for (some) other custom-nodes.

In practicality that means if you remove these connections you won't be able to control large aspects of the workflow anymore, mainly the detailers. Unless you remove the connection in the detailers too and manually change your scheduler there, you won't be able to use them correctly.

The easiest solution would be to deactivate/uninstall the RES4LYF node or ask the developer of the custom node to fix this problem by globally overwriting the scheduler selection for all other loaded nodes.

no problem @clawsofdeath :D
If you're not intending to use the detailers you can probably just continue generating like this ^^
I'll see if i can move the node developer of either the RES4LYF node or the Detailer node to implement a change to fix this in the near future! 🫡

@LemurT yeah i just found that out today too in my conversation here :D

I'll add it to the FAQ!

Heya @LemurT,
It's not a stupid question and i also have it disabled since it also annoys me 😂
To not further trigger your poor dog you can go into the settings in comfyui -> Image Filter(left sidebar) and then deactivate "Play sound when activating".

The sound is triggered by the image filter that lets you choose which image to continue the process with when you're doing a batch of higher than 1 ^^
You can either disable the sound like i mentioned above or simply disable the "Image Selection" in the Process-Control Group if you don't want to use that functionality altogether!

Hey @kitblasters339,
thank you for the kind words! I'm glad you're enjoying the workflows and the work that goes into them 🥰

You can definitely solve this issue with either the Face Detailer or the Manual Inpainting, but only if your checkpoint knows the face. Then you can simply mask the face, prompt what character you want to replace it by and then play around with the denoise to get it right.
You can also load a lora in for the character to help along the way.

But currently there is sadly no intention to implement InstantID or ReActor as of right now :/
Mostly because of their NSFW restrictions and because of some dependency meddling that can lead to many problems to most users. Also most users don't really know how to create the face models and it's not really an easy thing to do in a one-step workflow like this :/
So sadly right now and in the forseeable future the only way to achieve these results would be with a checkpoint and/or lora that is trained on the subject you want to display the face of.
You can also use the facedetailer to automatically mask the face area and use the txt2img workflow just to detail the face and prompt which face to display instead of the original 🫡

@kitblasters339 i have tried IPAdapter with Face in the past and for me at least it has not worked well. As long as the character's facial position and expression are the same it usually works okay'ish but as soon as that changes it's over.

I usually just rely on the prompts to be honest.
SDXL based checkpoints like illustrious, pony and noobai aren't really great at keeping the likeness of not well known fictional characters.
With what you're describing i'd usually try to use a lora for the character and if one doesn't exist for it i'd train it. With increased learning rate you can push down on the steps required in total and already achieve very high likeness with barely 1000 steps of training. On my 4090 that'd take around 40 minutes, which i'd be willing to sacrifice if i want to generate that character regularly.
It also helps that you can "overtrain" the character a bit since you'd only use it for inpainting in this case which is easy if the lora is only trained on the character and no other concept.

If you really want to implement instantid to the workflow it shouldn't be super complicated, just follow this node setup and put it before the model/prompts. You can do that by folding out the "Sampling Process"-Node in the "Sampling"-Group at the bottom of the workflow.
This one.

From the ingoing connections you remove the "model (after loras&ipa)" connection and put it in the "Apply InstantID"-Node as the model. Then you'll put the output of the "Apply InstantID" Model output back into the "Sampling Process"-Subgraph as the model(after loras & ipa).
From the outgoing connections you take the positive prompt & negative prompt and put them into the "Apply InstantID". Then put the output positive & negative from the "Apply InstantID"-Node into the "Sampling Process" as the positive and negative.

With this you'll have implemented InstantID to the workflow the same way you'd do it normally.

@kitblasters339 no worries my friend :)
I hope you can do it! i would've done it myself but i was too lazy to install the instantid nodes :D

I usually always train my loras with kohya-ss, but it's somewhat complicated and not really necessarily in a good way. I have heard great things from OneTrainer so you should definitely give it a try if you want to start trying to train a lora.
It definitely does a better job at end-user experience for beginners!

@kitblasters339 i'm glad you got it to work!
Yeah like i mentioned earlier, putting instantid into a workflow like this, depending on the amount of actions, will take quite a bit of time - especially on the higher end of the resolutions.

If you regularly try to use that character, creating a lora is definitely the easiest way!

Hey @bruceflee,

can you check if you have all custom nodes up-to-date and then refresh your browser once and tell me if the error still occurs?

Node #133 in particular would be the steps selector in the "Sampler Settings"-Group.
If you're having an issue, rightlicking and selecting "Fix node (recreate)" might fix the problem.

@Persson yeah sometimes comfyui is a very unique piece of software... 😂
Always make sure to refresh your browser after installing any nodes, a hard refresh can also always help to reset the cache by pressing CTRL+Shift+R in most browsers 🫡

@Persson been there 🥲
Hope you'll enjoy the workflow and that the struggle was worth it! 😂

@Persson happy to hear :)
It's easiest to place the notification sound node before/after the save image node, since that's the last step no matter what options you choose.

I'd recommend the individual detailers over the all in one, the all in one detailers are great for very simple workfows but not so much for stuff like this ^^
If you want to use the all in one detailer you have to know the keywords this detailer was trained on, for example "nipple", "pussy" etc.
And then wenn you use the all in one model you have to open the detailer itself and at the bottom of the node it says "wildcard spec" which is a text input, there you'll have to write the objects it's supposed to find.
For example:

pussy, nipple

This way it'll mask every pussy and nipple it can find from the image - but obviously it'd still output them as one batch, so you'd still have to chain these :D

hey @dennyma2012961,

are you using the latest version? In the version 5.0 i have removed anything anywhere from the workflow because of issues it caused with newer comfy versions.

hey @generator1337,

have you refreshed your browser once after doing all of that? Maybe also try a hard refresh by doing : CTRL+Shift+R

from the list of the input type i can see that you have the beta57 scheduler which is part of either of the nodes you mentioned. So it seems that something is still persisting these schedulers in your comfy installation.

I'd check the custom_nodes folder and make sure that they are a 100% gone and delete them otherwise, sometimes Comfy Manager can be a bit buggy and not correctly show all installed nodes. Make sure to turn off comfy afterwards and start it again and then try hard refreshing with the key combination mentioned above.

@generator1337 glad that worked!
The RES4LYF node can be a real pain in the ass :'D

Hey @eadespoker735,

haven't had anyone having this problem so far 😂
It's "only" 17 and a lot of them are nodes basically everyone has installed like KJNodes, Impact-Pack, Custom-Scripts and rgthree. But i'm currently working on reducing the custom nodes needed since in the last year comfy itself delivered some of the functionalities itself! :)

Are there any specific custom nodes you're having issues with? Maybe there are incompatibilities with custom-nodes you already have installed.
Is your comfy version on the newest current version?

I do not have a script to only install the custom nodes, but i do have a script to do a fresh clean install of comfyui with all required custom nodes that are needed for the workflow + the newest version of the workflow itself.
You can find it here(the windows-nvidia.bat):
https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/tree/main

@MiddlingMaker Ah that sucks :/
Usually you can fix it by doing a hard cache refresh (CTRL+Shift+R) and then reloading the workflow by dragging the file into comfy again. But yeah, comfy caching is hard and aggressive (should automatically refresh after custom node installation imo, but whatever 🫣)

@MiddlingMaker happy to hear! ♥️
Yes - the manual inpainting uses the whole model structure of the default workflow.
So you can change the model, loras and inject IPAdapter by activating their functionality's and selecting the models in the left part of the workflow.
This is done so you have one streamlined process where you might generate an image that you're happy with apart from a few details and then simply pop on the "Manual Inpainting", draw a mask and can edit the image straight away without having to add/edit the models, loras and/or ipadapter manually in the inpainting area again.

So every lora you add on the left side in the "LoRA's"-Group will be used for the whole process including the Manual Inpainting part! Hope that helps 🥰

Hey @OrchidAI,
i know it can be a bit overwhelming at first - i tried delivering all the options and knobs you might want to use during image generation with the intent to delivering everything A1111 had to offer for image generation in a single workflow.

If you want to prompt a character with blue eyes you should put it in the main prompt, only that is used during image generation.
If you do want to use the detailers for face and eyes you can also add it there as it'll help the detailer to guide the eye color. It's not really required or necessary when you're using low denoise values like 0.1 or 0.3, but if you use a higher denoise value it might change the eye color by itself and then it can help to explicitly prompt it in the detailer prompts.
You can also use it to change the eye color later on when using the img2img workflow for example. The order is always: Positive Prompt > Generates Image > Detailer > Uses Detailer prompts as help for generation of the masked area (if activated)

Hope that helps! :)

Thank you so much for your kind words @Valyria! ♥️
Always super happy to know people use the workflow and are enjoying it 🥰

Thanks a lot @MiddlingMaker! ♥️🫡

Thank you @MoHaXkm,
I'm very happy to hear that! ♥️

If you don't want to go through the hassle of hopping into your venv and doing all that manually, you can also go to Comfy Manager and find the node there and click "Try fix", that'll queue the installation of dependencys for your next start so you can simply click "restart" afterwards and upon restarting comfy will install necessary requirements automatically!

Hey @MiddlingMaker,

the IMG2IMG Workflow does not support using outside/your own ControlNet skeletons.
You do not not have to invert, shuffle or manually do anything.

If you're talking about the IMG2IMG workflow and the IMG2IMG Transfer-Functionality you can find the subgraph here at the bottom and in there you can find the ControlNet-Node.

The Pre-Processor right before that is what creates the pose image.

I have never heard of someone having to invert the Red and Blue channels of OpenPose Skeletons for the ControlNet to work. Have you tried simply increasing the strength(accuracy)?
If your prompt heavily contradicts your controlnet, with a default strength of 0.55 you will face problems locking that pose in. Prompts heavily matter which is why it has an inbuilt interrogation.

You also shouldn't have to invert the R&B Channels of your OpenPose Skeletons for the TXT2IMG either - are you sure you're using all the correct nodes and have all of them up-to-date with their dependencys installed?

Hey @Linpeak,
you do seem to be having issues with the mixlab node as seen here:
"D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/photoswipe-lightbox.esm.min.js
D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/photoswipe.min.css
D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/pickr.min.js
D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/classic.min.css
D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/juxtapose.css
D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/model-viewer.min.js
D:\AIsoftware\ComfyUI-aki\ComfyUI\custom_nodes\comfyui-mixlab-nodes\webApp\lib/juxtapose.min.js"

But that node-pack is not part of the workflow.
It might be that it has simply broke the javascript of your comfyui after updating and that's why some of the other nodes are having issues displaying their javascript.
I'd recommend uninstalling the mixlab-nodes if you don't need them for anything else.
You might have mistaken the mixlab-nodes for the actual nodepack that is required for this workflow:
https://github.com/Smirnov75/ComfyUI-mxToolkit

Make sure you have the correct one installed.

Hey @boz255,
sadly having this many functions in one workflow result in a lot of wires 😂
Not much i can do about that sadly.

I have just set up a clean instance with the newest comfyui and all custom nodes required completely up-to-date and cannot reproduce this error at all.
Could you tell me which functionalitys you have activated?

My guess is that the "Detect Corrupt Workflows" functionality of rgthree did not fix anything but instead destroyed it.
These "auto-find" functions to check if workflows have issues simply don't work very well with subgraphs and bypassed nodes. It'll mark nodes as missing their input or having loops because their predecessors are bypassed, even if that simply means that they'd forward their input.

My recommendation would be to open the rgthree settings and deactivate "Deactivate corrupt workflows".
You can also deactivate the "Validate workflows" function in the comfy settings if you're bothered by the toast message that there are missing links.

Again - the workflow works perfectly fine with the newest comfy and plugin versions, these tools are simply not meant for workflows of this size and don't do a good job beyond simple relations of a few nodes.

edit: just checked and did the "Fix in place" myself and it still worked afterswards, might depend on the functions you have activated though. If you let me know which functions you tried to use i can check if the fixed version broke something that works in the non-fixed version.

@boz255 Yeah i totally get it, working through other peoples workflows is already a pain in the ass, going through one this big is diabolical. Hard to figured out what‘s needed for what.

It‘s fine if the output comes out of a subgraph, into a node and back into the subgraph - since the subgraph itself doesn‘t count as a single node entity. So this shouldn‘t cause a loop error unless something is broken. Have you installed the missing nodes through the comfy „install missing nodes“ function?

I had someone today who ran into an issue because the comfy functionality actually downloaded another custom node called Comfy_Dummy_Nodes that overwrote another node from advanced controlnet and therefore causing an issue. If you can make sure you don‘t have that one installed.

Lastly you could try using the 5.1 that i‘m currently working on. You can find it here even if it‘s not done yet, it already has some improvements including fixes that should not cause the „fix workflow“ errors anymore. There are still 1-2 todo‘s before i‘ll publish it on civitai but it‘s already good to use. Make sure that all the required custom nodes are up to date through the comfy manager and let me know if the 5.1 works for you.

Hey @OliverAntheia,
glad to have you back using the workflow! :D

I'm still working on an update for the guide since the 5.0 which now includes inpainting has added a lot of extra content ^^
Once that's done i'll include the recommendation for the face detector models, i have already used them myself and i think they are the best ones out there! Ty for the recommendation :>

Hey @RTSMT,

the portable version is usually more prone to having some issues - this seems like a usual "can't find the node" problem.
It might be that the upscale by factor node is hard cached somehow in your comfyui.
I'd recommend going into your custom_nodes folder and manually deleting the whole vslinx-nodes folder.
Then pull the node through git here instead of using the comfy manager with this:
git pull https://github.com/vslinx/ComfyUI-vslinx-nodes.git

this garantuees you have the most recent version 100%

Then you restart your comfyui and then do a hard refresh by pressing CTRL+Shift+R in your browser where you have comfy open - this will reset comfy's cache.

If that doesn't work please try the 5.1 of the workflow that is currently still work in progress to see if that works. You can get it here:
https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/blob/main/workflows/TXT2IMG/v5.1/TXT2IMG-ADetailer-v5.1-vslinx.json

@RTSMT sorry man 😂
The portable version is sadly really really ass.
It might not come with all necessary dependencys :/

I'd always recommend pulling comfy via git to make it as easily updateble as possible.
It's also basically still a portable version since it installs all dependencys locally in it's own virtual environment.

Heya @kitblasters339,

in my opinion the best SDXL checkpoint for realism is 100% Lustify. It should work perfectly with the automatically generated prompts from the IMG2IMG Transfer.
If you're going to use the manual inpainting area i recommend using the Lustify Inpaint Checkpoint.
Both of them are in my opinion the best you can get out of any SDXL based realism checkpoint.

Hey @Kaymart,
You should check the FAQ section here.

The "workflow" in the name is always automatically prepended when throwing this error, so the actual thing that cannot be found is the "ControlNet Models"-Subgraph in the ControlNet Group.

There's only 2 nodes inside, the Advanced-ControlNet Custom-Node and the comfyui_controlnet_aux Custom-Node.

So there is one of them with which you're having issues with.
If you have used the default comfy function to install all missing custom nodes it's also very very likely you have downloaded an additional node that'd break things.

There is a node called Comfy_Dummy_Nodes that get's automatically installed because it has the same ID as the advanced controlnet. It breaks the Advanced ControlNet node so you have to make sure to uninstall if you have it.

Hey @PeachTheBunny,

are you using the TXT2IMG or the IMG2IMG workflow?
If you're using the IMG2IMG workflow it might be worth for you to check this guide:
https://civitai.com/articles/15480

If you want to just check what you have to do for a specific scenario you can go down to the scenario section here:
https://civitai.com/articles/15480#scenarios

@PeachTheBunny what workflow did you download? Or what is the file name?
It usually says either TXT2IMG or IMG2IMG in the filename, here on the page is a TXT2IMG and IMG2IMG workflow for every version. But according to your comment you're probably trying the txt2img one.

I think you're looking at the Image feed and not at the actual result, the image feed just shows you all images you have created during the generation of the workflow, including previews - not just the image that you have saved. You can change this in the settings of comfyui to prevent your image feed being filled with previews.

Since you're not getting an actual final result i'd assume that you're probably getting an error, not every failed execution will throw a visible error in comfyui itself, most issues will just silently show in the console.
So could you check the console where you start comfyui and post the error that you have there after executing the workflow?

Also: You don't really need the files for face, hands etc. if you don't want to use the detailers. They should not throw an error as long as you don't use the detailers.
If i had to guess you probably need to do two things if you haven't already:

- Select a valid model in the "Refiner Checkpoint" in the second row. Even if you don't want to use it, you sadly still have to select a model here, it's comfyui's behaviour :/

- Make sure you see an image in the "Manual Inpainting"-Area, if there is no image in the "Load last generated image"-Node. If there is not that'd mean you have no image directly in your comfyui output folder. Sadly it's a comfy core node that has a hard check on wether you have an image here and it'll fail if you don't. It's annoying but usually only an issue if you're using a fresh comfy installation.

@PeachTheBunny then i assume that you have the seed set to „fixed“.

The seed is what decides what the image output is based on your prompt and other settings. By keeping the seed fixed you are able to re-create the image even after generating it once.

If you want to create a new image every time you have to set the „control after generate“ to „randomize“ in the „seed“-node inside the „Sampler Settings“-Group

@PeachTheBunny glad to be of help! 🥰

Hope you‘ll enjoy the workflow! 🫶🏽

Hey @arimintzu,

sorry today i was a bit late with checking civitai :D

Yeah if you install all missing nodes simply through the „install all missing nodes“-feature, comfy will install every node pack that has the same name for the node…

Sadly the dummy developer was a dummy and used the same name as the very popular node pack :D

Does everything work now after uninstalling the wrong conflicting node pack?

Hey @alesanand97,

this is not a simple or easy workflow - for that there's ton's of easier options here on civitai ^^
This workflow is an All-in-One solution, if you have used Automatic1111's WebUI before and you want the same feature-set, this workflow is most likely your best shot.
It has a lot of knobs and values that make it easy to use for an end-user without having to have a deep understanding of comfyui itself. This is who this workflow is aimed at.
If you don't have any idea what controlnet, ipadapter, hires fix, inpainting, adetailer or refiner is - this workflow is probably a bit too much functionality to grasp at the beginning.

There is no youtube video but there is a guide you can find in the description (or here).
This is for the img2img varient but it mostly explains everything the workflow has to offer and is pretty similar to the txt2img version. The guide is 2-3 versions behind and doesnt include the newest manual inpainting area.

There are also notes inside the workflow that explain the functionality of all user-relevant nodes and the groups directly inside the workflow.

Hey @arimintzu,

in the "Load ControlNet Pose"-Node you can upload a pose that's already prepared for ControlNet, like you can see here.

If you don't already have a Pose for the specific ControlNet, you first need to create a pose image. You can do that by just uploading any image that you want and turn it into a pose by selecting a preprocessor, uploading the image to "Pre-Process Image to Pose" and then activating the preprocessor function together with the controlnet function.

This will turn your image into a pose, which you can see a preview of in the Debug section, and the workflow will then use that pose to create your image.

Hey @MindAddiction,
no worries - there are no stupid questions :D

This is the first time i have heard of this issue, but i have found similar problems on their github repo. What GPU do you have? Is it a non-nvidia or a 50 series?

Seemingly there are fixes for this problem that have yet to be merged but the node hasnt received an update in 5 months, so my hopes are low on that regard.

BUT i have been developing the 5.1 version of the workflow where i mainly make the workflow more stable and remove a lot of unnecessary custom nodes to make the requirements easier.
Crystools was one of these custom_nodes that i have removed from the workflow.
You can find the TXT2IMG 5.1 here and the IMG2IMG 5.1 here.

They are basically done already, there's just one bug i want to iron out in the img2img that's very nieche and probably not really problematic for most people and for the txt2img i'm just waiting for 2 merge-requests being merged in some custom_nodes since they'd make the workflow more performant.
But they are already good to use and should help you fix that problem since they don't use Crystools anymore, so you can delete that node from yours.

@MindAddiction no need to thank me, usually i'm more active and answer faster :D
But the last few days have been a bit busy.

Weird, maybe you don't have cuda specifically installed, but it should not break the nodes.
I guess i'll roll out the 5.1 in the following days without waiting for the merge-requests since this'll probably happen more often now.
My best guess is that ComfyUI doesn't include CUDA as a dependency by default anymore and Crystools references it ^^

Glad the workflows are helping you!
Sorry that the workflow seemingly got slower :/ My best guess is that they have changed something with the memory allocation. That's sadly not something i have a lot of control over :(

@MindAddiction Yeah i also sometimes have this issue.
Sometimes when comfyui updates i'm running into super long VAE Decode processes, when it'd usually take half a second it all of a sudden takes 40-50 seconds. And that's on a 4090.

And then after another comfyui update it works fine all of a sudden again. Sometimes it's worth not upgrading haha :D

Thank you! Hope your issue get's fixed soon and i hope that the workflow will help you 🫡♥️

@MindAddiction 

PS: Was running into the same issue again today, i think the latest comfyui update changed their smart memory logic and it now offloads VAE Decoding onto your CPU instead of GPU when it thinks that you don't have enough vram available.

For me it only happened during the detailers - so you can simply open the Detailer and select "Tiled Encode" and "Tiled Decode" in the Detailer settings, it then does it automatically in smaller chunks and doesn't offload to the CPU.

You can also disable the smart-memory shit globally for comfy by using "--disable-smart-memory" in the starting arguments of your comfy start. You can also force your comfy to always do everything your gpu only by using "--gpu-only".
Neither are super good if you're doing very vram heavy tasks that have to offload to CPU/RAM, like generating videos or something - but it'd help a lot with image generation where you know you have enough vram and comfy is just being dumb with it's smart memory management.

Heya @155T,

thanks for the report.
Seemingly the newest comfyui update has broken quite a lot of things... 🥲
Not just with my workflow but inside the comfyui frontend in general.
Load image nodes work awfully right now and will empty the preview and output when you switch images or workflow tabs...
I am right now developing my own "load last generated image"-node to use instead of the comfy one, since it has been a painpoint in the past and i think i can do better 😂
I'll release a new version later today, thanks again for bringing this to my attention!

Hey @last27683599,

thanks for the report.
Seemingly the newest comfyui update has broken quite a lot of things... 🥲
Not just with my workflow but inside the comfyui frontend in general.
Load image nodes work awfully right now and will empty the preview and output when you switch images or workflow tabs...
I am right now developing my own "load last generated image"-node to use instead of the comfy one, since it has been a painpoint in the past and i think i can do better 😂
I'll release a new version later today, thanks again for bringing this to my attention!