ComfyUI Detailer/ADetailer Workflow

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.

@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.

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

@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! 🫡♥️

@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:

https://github.com/ltdrdata/ComfyUI-Impact-Pack/pull/1115

@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 🫡

@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.

@PantyMilk More buttons = better 😂
Thank you so much for the buzz! ♥️

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.

@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.

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.

@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.

@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.

@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 ^^

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?

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 🥰♥️

@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

Hey @bacsvord123464,

what is the console saying?

@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.

@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.

@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 🫡♥️

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 @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.

@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.

@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.

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 :)

@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.

@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 :/

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 :)

@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.

@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)

@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

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! 🥰

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 😎

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 ♥️

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".

@Taimanin001 glad it's working now! and sorry you had to reinstall comfy :/
It sometimes can be a hassle!

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!

ayyy @TootyPenguin i'm happy to hear that!
Thanks for the feedback 🥰

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

@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.

@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?

@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.

@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!

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.

@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!

@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.

@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 🥰

@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

@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! 🥰

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.

Hey @harlemrayx,

please check out the error in the FAQ of the guide.
Hope this helps 🥰

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 (✿◡‿◡)

@chloeChu very happy to hear!
Glad you're enjoying the workflow ^-^

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?

@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)"

@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.

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

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 ^-^

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 🙏

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 🥰

@blackjuker you're buzzing man! 😂♥️ thank you a ton!

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 🫣

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.

i‘m glad if that helped you @HazelYK! ❤️

Thanks a lot for the buzz donation! 🫶🏽

Heya @CplGreedz,

first i wouldn't try to manually bypass the parts, that might destroy some of the functions and switches that are required for switching between the outputs. So only use the switches to activate/deactivate functions 🫡

Do you have any messages in the console where comfy is running?
Could you provide me with a screenshot through imgur or google drive or something?

When initially loading the workflow the only things that are required to be filled out by you, should be the checkpoint, vae and upscale model. From then on out you should be able to run the workflow.

There might be 1 or 2 nodes shown as "running" inside the inpaint part, but these are lazy switches that switch the output based on the switches you have activated. They just serve to forward the image for the previews and do not actually mean that the inpaint area is currently active.

hey @CplGreedz,

noo need to put yourself down - we all start somewhere and Comfy is sadly not the easiest tool to get into ^^

The two errors i can see come from 2 things and are really easy to fix.
The first: "Checkpoint failure" comes from the seconds row -> second group called "Refiner".
When executing the workflow it tries to check for the checkpoint and cant find it in your list.
Just change it to any of your checkpoints like the prefectIllustriousXL_v10 for example - no need to activate or use it, you just need to have an existing checkpoint here.
This is where you have to change it.

The second: "Output images will be put here" means that the "Load last image from output"-Node cannot find any images in your output directory - most likely because you have not yet created any images and therefore your directory is empty.
You can just click on the "choose file to upload"-Button in the Inpainting area to select any picture - no matter where, to fix this issue. Once you have generated 1 image this won't appear again :)
This is where you have to change it.

Let me know if everything works now or if you still need help 🫡♥️

PS: The second image shows warnings from comfyui, the newest frontend version now has something called workflow validation which is not really functional yet and still in the beta phase of it. You can safely ignore them since it's just comfy being stupid 🫣😂

hey @CplGreedz,

Glad that it's working now!
It should only be slower if you're using a lot of resources on your computer, if you're having a black image instead of a generated one it's most likely something with the workflow itself :o

Let me know if you get black images again and if there are any errors in the console.
If you're using the detailers and you're getting black images in the previews it's most likely just that the ultralytics model couldn't find what it's trained to find in an image. But if you're getting a black image in one of the two preview fields let me know!

@xCraftyx this workflow is meant to be run locally via comfyui only.

I have gotten invites for tensor.art in the past but because it‘s such a closed off system i don‘t like to have my workflow published there since i have virtually no way to debug if anything breaks :/

@RussianMan_BRUTAL that sounds unlikely unless you're on a very old comfy version or you're missing custom nodes.
If you copy and paste your errors i might be able to help you figure out what the issue is 🫡🙏

i‘ll gladly help @RussianMan_BRUTAL :D

Just send me the full error log of your console when executing the workflow and i‘ll help you as good as i can! 🫶🏽

Hey @RussianMan_BRUTAL,

i don't really give tutorials on what IPAdapter or ControlNet are since there are a lot of good tutorials out there that probably explain it way better than i do.

You can however find brief explanations in my IMG2IMG guide about what these are.
Here for IPAdapter and here for ControlNet.
There you can also find an explanation on how to use the IPAdapter, Otherwise there are explanations on how to use the IPAdapter and the ControlNet inside the Explanation node to the left called "How to use".

I tried to make the process as streamlined as possible and the workflow is mostly meant for people who already know what all these things are and only want to use them. It's probably not a good workflow for someone trying to learn some additional things.
Don't use/activate the Manual Inpainting part of the workflow if you don't actually want to inpaint anything, for normal text generation you will not need this area.
I can only recommend reading the notes inside the workflow again for ONLY the functions you want to use and only activating the functions you want to use and know what they do.
If you wildly activate a lot of options and don't know how they work or don't provide images for the process it's normal that some things might not work.

I hope that helps 🫡

@RussianMan_BRUTAL sometimes your pc doesn't have the permissions to access the file from outside the comfy folder. You can then just open the .csv file with editor, copy all the content and paste it into the field in the settings instead of setting the file path.

You can then click save and it should work! 🥰

@RussianMan_BRUTAL If you have a question about the functionality of the workflow you can always ask and i'll always try to help :D

If it's more about general stuff like what controlnet, ipadapter, inpainting, refiner etc. is, it's probably better to google or ask chatgpt since it'll solve your issues faster :D

@RussianMan_BRUTAL both of these models are ultralytics segment (seg/segm) models.
So they both belong into the ultralytics/segm folder.

They are NOT loras, i think you don't quite understand what they are doing or what they are there for. They are detection models only there to recognize and mask areas it is trained on, in this example the face.

So all it does is mask the face area so that the actual detailer can then re-paint them at a higher resolution.

You can also find that information in the Guide where it says that:
"Detailer (ADetailer)

The Detailer (often called ADetailer) is used to refine specific parts of an image after the initial generation. It detects body parts or facial features—like eyes, hands, or mouth—using object detection (usually with Ultralytics), and then re-renders those areas at higher quality using targeted prompts and LoRAs. It’s great for fixing anatomy issues or enhancing details like eye color or hand shape."

Please always check the guide or notes inside the workflow first, this way you don't have to wait hours or days for an answer and can help yourself.

In the Files section you can also see an explanation of where they're supposed to go.

hey @rivverswimmer2572,

yeah this problem is listed in the FAQ in the description ^^
It's sadly not possible for me to create a workaround anymore since comfyui changed some things and now the invasive RES4LYF custom_node just messes up things with other nodes like the detailers :/

Thank you! ♥️🥰 really appreciate it

Hey @xenith617,

this sounds weird - normally the controlnet pose should never even end up anywhere close to the output/preview nodes apart from the preview in the debug section.
Are you maybe using the Image feed to check for results? The image feed will show every single preview as a "generated image" unless you change the settings of comfy accordingly so that they only show saved images. Or is the ControlNet Pose actually ending up in the Output folder?

Can you tell me if you have any errors in the console of your comfy? Maybe there is one file missing.
Usually upon loading you'd only have to change a few models for the workflow to run.
The Checkpoint, VAE, Upscale Model and the Checkpoint selected in the Refiner(even if you don't use it, comfy requires a valid checkpoint here).
Let me know if you have filled all of these out and what your console is saying when trying to generate and image 🫡♥️
And don't ever be shy to ask! I'd rather have someone enjoy my workflow than despair because i didn't make something as clear or convenient as intended ^^

@xenith617 yeah sadly there is nothing i can do to fix that :/
Previously in comfy i was able to use some converters to convert present to expected schedulers, but due to some recent changes on Comfy this is no longer possible.
This issue sadly happens a lot and is also listed in my FAQ section.

I hope one day comfy will either allow us to do non-strict passing of values or the custom_node developer of RES4LYF will do a different implementation.

Hey @last27683599,

The image is only ever saved after appearing in the final preview.
If there is no image appearing in the preview then there is also non saved and if it's not appearing in the preview that means that there was an error during generation.

It's good you're not using node 2.0 since they won't work due to incompatibility with rgthree-nodes, but you should have an error in your console that could give us an idea on what's broken.

hey @last27683599,

that error report is already really helpful!
First of all you're getting some warnings like "[ComfyUI-Manager] Skipped fixing the 'comfyui-frontend-package' dependency because the ComfyUI is outdated." and later on the "comfyui-frontend-package not found in requirements.txt" which most likely means that your frontend package is (very) outdated. That means while your comfy(backend) does have most of the functionality, your frontend cant display nor pass the correct values to the actual technical part of the image generation.

Sadly i saw that in your error log:
"### ComfyUI Revision: UNKNOWN (The currently installed ComfyUI is not a Git repository)"
Which means you can't just simply pull the new frontend because you most likely downloaded comfy as a zip file and then unpacked it somewhere on your PC instead of "pulling" it with something called "git" which serves to keep version control of some you download via git and enables you to update it through commands.

From the looks of it you also probably installed all comfy dependencys globally, which can be a hassle since globally installed packages might end up conflicting with other software on your computer which might result in incompatibilities with other software that might also use and/or update it.
The best practice is to start comfy in a venv which means a virtual environment, that helps to keep everything installed necessary for comfy to ONLY be installed inside the comfy directory. So that nothing is really globally installed on your computer and only comfy uses these dependencys. With this you can easily manage the versions of everything installed and it's also what comfy suggests you should do in their installation guide.

The last error comes from you not having a "example.png" file in your input folder inside comfy.
Usually comfy always comes with one image inside the input folder as somewhat of an example.
This image is also visible in my workflow images and it's the little drawing of this one figure with the big yellow hair/ears.
If that image isn't present the workflow will fail since it has to have an image to display in the default "Load image node" from Comfy. So you can either select a new image in every "load image"-node inside the workflow OR you place one image into your /comfyui/input folder and call it "example.png".
Note: you also have to upload at least 1 Image in the Inpaint Area into the "Load last generated image"-Node. But only if you have never generated an image with comfy before and there is no preview of another image in that node.

I know this is all a lot and i hope i explained everything as good as possible :D
But sadly your comfyui installation is a little bit fucked/corrupted by the way you've downloaded and installed it. If you don't want to do all these steps and want to be future-proof with your comfy installation you could use the install script that i use myself to re-install my comfy installation including all dependencys/custom_nodes needed for this workflow.
I use it myself whenever i test a new version before release and a few others have also already used it to install comfy correctly.

You can find it here:
https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/tree/main

You can also open the file and check what it does to make sure nothing fishy is going on, it basically just downloads the newest comfy+frontend version and then installs all dependencys only in the comfy directory and gives you the option to also download the controlnet models etc.
At the end it lets you create a desktop shortcut which will ALWAYS start comfy in it's own venv so that it never installs anything outside of it's own directory.
You can try it if you want to - but no pressure, the other steps above will probably already fix your issue unless the frontend is not updateable at all.
If you want to use the script just download the "windows-nvidia.bat" from the link above and put it where you want comfy installed and then just execute it. It'll install comfy correctly right then and there.

@last27683599 nice i'm glad it worked! ♥️
Yeah the portable version would also be wrong ^^ it's best to pull the actual directory via git, but if comfy updated just fine from the unpacked zip, it's also fine - glad it's working now! 🥰

Hope you'll enjoy the workflow and let me know if something's not working!

Hey @RussianMan_BRUTAL,

you can also find that Information in the notes of the workflow itself:
"ControlNet (Optional)

You can either provide a full image in the "Pre-Process image to Pose" image selector or a ControlNet pose image in the "Load ControlNet Pose" image selector at the bottom.
The ControlNet image has to match the ControlNet model in the node inside of the "ControlNet Model"-Node at the top.
If you want to imitate the pose of an image and have selected a image in the "Pre-Process image to Pose"-Node inside the "ControlNet"-Group you have to enable the "Use Pre-Processor"-Node in the "IPAdapter / ControlNet"-Group.
After selecting a pre-processing model at the bottom of the "Model Backend"-Group you can see a preview of the generated/used pose in the "ControlNet Pose Preview"-Node inside the "Debug"-Group.

You can find an in-depth guide about what ControlNet is, what it does and how each model works in this guide. (It's for A1111WebUI but you'll understand)"

I recommend reading the guide that i linked there to understand what ControlNet even does.

In short you can either upload a controlnet pose image, where the image is already pre-processed like these - OR you just provide an Image that you want to copy the pose from (Pre-Process Image) and activate the pre-processing in the controlnet group.

hey @whateverr,

You don't always get an error thrown from comfy, is there something in the console?
There should be an error in your console where comfy is running if something fails and the workflow just stops without generating anything.

hey @tise4ka406,

yeah that is sadly an issue with a comfy update that broke some things with some custom nodes :/
it really really sucks and i have already brought to their attention here:
https://github.com/Comfy-Org/ComfyUI_frontend/issues/6633

This issue has been fixed with this merge-request thanks to a fellow developer in github:
https://github.com/Comfy-Org/ComfyUI_frontend/pull/7154

It's already merged so the fix is already approved, it's just now waiting to be bundles into the next release and will most likely be included in the next comfy update! 🫡♥️

hey @Rvage,

yeah that's what i meant with bundled into the next (frontend) release :D my bad for badly wording it ^^
Thanks a lot for your fix in the github repo! ♥️

Hey @SugoiDess,

if there is no image saved to the output folder it‘d mean that the process didn‘t fully finish. The output images will have all the metadata including used loras.

Can you tell me what errors you have in your console when running the workflow?

Hey @SugoiDess,
these are a lot of errors upon startup for a new installation - damn!
But they are mostly from the GGUF-FantasyTalking node so it should be fine.

The workflow can't run because you're missing a lot of the models required for it or rather don't have them in your models folder.

This error:
* SAMLoader 9:
- Value not in list: model_name: 'sam_vit_b_01ec64.pth' not in []
Can be fixed with this here.

All the other errors are from you not having selected a Checkpoint/VAE/Detection Model.
The workflow you currently have open already has some models selected but it seems that your checkpoint folder does not have the "betterDaysIllustriousXL_illuXLV0341BKEDVAE.safetensors" checkpoint (it seems to be empty in fact).
You don't have a vae selected, the default sdxl one is not in your folder - the only one in your vae folder is the pixel_space vae.
You're also missing a face detection model while having the face detailer activated, you've selected the "segm/Anzhc Face seg 1024 v2 y8n.pt"-File but there are no files in your ultralytics folder.

In general i wouldn't recommend using portable versions, the sam_vit_b model for example should already be included in the default comfy installation.
The Tavriv1's install script also seems to also be based on the portable comfy version.
If you want to do a clean full "install" of comfyui in it's own virtual environment i could recommend my install script that installs everything automatically that is required for this workflow (except the models of course). It'd install comfy in it's base version wherever you execute it and will run in it's own venv so nothing get's installed globally which helps a lot with dependencies for custom nodes.
You can find the windows-nvidia.bat here if you're interested. But i'd also understand if you don't want to do a new comfy installation again, so the suggestions above should already fix your problems.

@SugoiDess you are trying to install a new comfyui where the old one already is.
So the install script is basically skipping everything.
That's why it's saying this everywhere:
Cloning custom nodes...
- ComfyUI-Impact-Pack already exists. Skipping.
Installing requirements for ComfyUI-Impact-Pack...

I assume the desktop shortcut couldn't be created because of some file permission issues with your computer.
You should install it with the script somewhere else than your current comfy installation, you can just rename the old ComfyUI folder to something else like ComfyUI2 to make the actual script run without issues.

After that you shouldn't have any issues.
It's also always important to refresh the page once after updating/installing anything, because comfyui keeps all old things in it's cache. Make sure to refresh the page and then load the downloaded workflow again, not your old edited version.

Lastly since it didn't generate a desktop shortcut for you because of file permission issues i recommend creating a new one where you start ComfyUI in a venv instead of just normally.
So create a shortcut on your desktop that does this:
@echo off

cd /d "G:\\AI\ComfyUI\"

call "G:\\AI\ComfyUI\venv\Scripts\activate.bat"

python main.py

pause

@SugoiDess Yeeeey! :D

Yeah for the last generated image node there always has to be at least 1 image in the output folder or at least 1 selected ^^ it‘s a bit annoying and i hope comfy will change the behaviour of the node soon!

Hope you‘ll enjoy the workflow ❤️

Hey @Unhing3d,
well that can be or good or bad thing! :D but i'll take it as a compliment for now 😎
This problem is relatively easy to fix, you can read the FAQ entry for that here.

Just a little mention - if you replace the lora loaders, the image saver will not save the metadata information for that ^^ So if you upload the images to civitai the loras wont be automatically added as meta info

@Unhing3d was just waiting for your question 😈😂
Glad it works now! (Hopefully) Enjoy the workflow! ♥️