ComfyUI Detailer/ADetailer Workflow

Hey @Feral_Signal_Studios,

Thanks so much for your Kind words! I really appreciate them 🥰

The custom nodes required for the workflow are all available through the comfy manager, they are mostly very popular custom nodes.

Just go into the Manager and click on „Install missing custom nodes“, just don‘t download the Dummy Nodes - They have a conflict with the naming of the Advanced ControlNet.

@Feral_Signal_Studios that's weird, i checked all of the nodes yesterday again after your comment and saw that comfyroll was able to be installed wihtout any issues through the comfy manager. Are you getting any specific errors?
Worst case you should be able to manually install the node by putting it into your comfyui/custom_nodes folder:
https://github.com/Suzie1/ComfyUI_Comfyroll_CustomNodes

(download either via git or by clicking on code -> download zip and then moving the contents)

@Feral_Signal_Studios that sucks :/
i have checked really quick and there are no known conflicting node packages, so unless you have installed some stuff manually and that is not available through the comfyui-manager browser.

If you ever quickly and easily want to do a clean install of comfy in it's newest version you could use my installation script that quite a few people have used until now.
It automatically downloads ComfyUI, all required nodes for this workflow incl. Comfy Manager and then gives you the option to also optionally download the controlnet models if you want and create a desktop shortcut at the end.
It'll execute wherever you open it and create a ComfyUI folder at that point.
Only requirements are python (3.12 is best) and git installed.
You can find it here:
https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/tree/main

Afterwards you could just copy your models folder over from your old installation.

@Feral_Signal_Studios happy to hear!
And i value all feedback equally, it might has a lot of downloads but most people commenting are usually because something isn't working :D so i always appreciate when people say thanks ^-^

@Feral_Signal_Studios Thanks, happy to hear you're enjoying the workflow!
It has become quite massive but i hope i'll get it to be more stable in the future 🫡
I might implement a 1-pass hires fix in the future if i hear anything about negative implications on the quality the way it's implemented right now, but until then it seems to work fine :P

Heya @last27683599,
no problem - thanks again for reporting the issue!
I don't always update my comfy the moment they release a new version, i might sometimes go days or weeks without noticing bugs in newer versions ^^

"Sadly" adding the body-specific detailers in the manual inpainting workflow wouldn't make sense, since you could simply re-run the workflow but with the eye/face detailer activated this time.
If you do not change the seed between generations, the image generated will be the same.
Right now re-executing the workflow sadly forces the image to be generated completely new again even if it looks the same, i opened a merge-request in the Impact-Pack over a month ago so that the image would be cached and not has to be generated again, but sadly the maintainer has not yet accepted the merge :/
But if you'd want you could apply the file change locally in your impact-pack node to solve this problem.

Additionally you could also just pass your generated image into the img2img workflow, deactivate the img2img transfer, upscaling etc. and only have the face/eye detailer activated.
This will just do an img2img where you only use detailers.

Also the manual inpainting should not result in lower quality outputs than the detailer itself, are you using the manual inpainting together with the hires fix option inside of it? I have not had worse results with the manual inpainter yet :o

@last27683599 phew that's a long read! hahaha

My thoughts:

- Denoise Strength of 1 is very very high, you basically do a full replacement of the content without giving the base any context to work with, so you 100% rely on the controlnet from that point onwards, just remember that
- ControlNet Strength of 1 is very high, it basically forces the character into that exact position with no leeway. I would not necessarily recommend it unless you use a specialized inpainting checkpoint.
I'd try to play around, i know you said that you tried 0.55 but that most likely didn't work together with the sampler. Strength of 0.52-0.55 should already copy the post near-identical, might have to play with the seed a bit, or increase strength to 0.6

- The Sampler DPM++ 2M is something i chose and activated by default for the most common user, most people would only want to change small things and for that the DPM++ 2M is definitely the best choice.
Different Samplers converge very differently, a good read for that topic would be this.
To explain simply some samplers spend more time in the earlier denoising steps, some spend more in the middle and some more in the end. DPM++ 2M is a multistep solver, meaning it doesn't decide how to add noise to the image to change it only based on the current point, but it uses information from previous steps. So if you're not starting on a blank canvas that means that it preserves structure way better, blends edges more naturally into unmasked regions, gives cleaner results at lower step counts and it's less likely to inject a lot of noise into the masked area.
For our images, especially anything involving denoise like inpainting, that means that they are closer to the original because they preserve more of the original. It's the best sampler to add some details, change smaller things like eye-color or clothes.
That's why i also changed the hires fix over to that sampler + scheduler(karras) because it preserves the original better, this is why now there's less artifacts and difference between the pre-hires fix and the hires-fix image. It's also the better alternative if you want to do detailing.
But that is also the reason why completely changing the content of the image would be harder with that or would mean you'd have to up the denoise level way higher than usually necessary, because it respects the form of the previous denoise step more than other samplers.

- Try changing the preprocessor for the controlnet and the controlnet model, maybe you're using one that is either not specific enough or too specific. Canny or MangaLine can force details from the original you might not want to copy because they are very harsh with their details even at lower controlnet strengths. Try MiDas Depth Map, i usually use that if i want to copy composition but not likeness of the character, it usually does a good job not moving too many of the original characters features over.
With simple poses like this you might also just be able to use OpenPose for your ControlNet, since it's quite the simple pose. OpenPose only copy's the skeleton and nothing about the appearance at all, so you can easily increase strength here.

- What Upscale Model are you using? I have had issues in the past with inpainting because i was using a shitty upscale model like Anime-Sharp or something.
These Upscalers remove a lot of details and are quite blurry - as long as you have hires fix on you won't notice it because it runs after that. But they really remove a lot of the images detail.
Using the inpainting you're doing something like this:
Generate Inpaint Area -> Upscaling -> hires fix -> Paste into original and blend

Depending on the strength of the denoise, the hires fix might not have a huge impact.
So you could try changing the upscale model and see if that improves sharpness.
I have moved from RealsESRGAN and AnimeSharp to "4x_NMKD_Siax_200k.pth", which i also set as the default and it really improved my image quality compared to earlier work of mine.
It will introduce some sharpness lines but the hires fix will remove these.
I have implemented a "Use different Upscale Model"-Section where you can replace the upscale model only for the inpainting, so you might want to play around here and check if a different upscale model could improve your result.

I hope any of these things were helpful :D!
And sorry for the long read - let me know if any of these things helped you!

@last27683599 
Here i did a replacement with your image.
I tweaked the settings a bit, but apart from that i loaded your image into the "Load last generated image"-Node and then rightclicked -> WD14 Tagger, copied the tags and added them to the inpaint prompt. I removed everything describing the hair and eyes, replaced them with the other characters and added the character tag at the beginning.
Then i drew the mask over the whole character and send it off.

This is the result, you should be able to download the image and drag it into your comfyui to load the workflow directly and see what settings for controlnet, denoise and sampler i used.

@last27683599 sorry man i just noticed that i had never answered 🥲

"So, I've been trying to emulate the workflow you've provided, but I've run into a slight issue. After right-clicking and selecting the WD14 tagger, nothing happens. Am I supposed to do something else after clicking on that option?"
- You can check your console, the first time you do that it'll usually take a little bit as it has to download the WD14 model for image interrogation. Then it usually just takes 4-5 seconds when doing that. A popup should open and your prompt should appear. It's basically the same the IMG2IMG Transfer does in the IMG2IMG Workflow.

"Generally, from my results, using controlnet and pre-processing, even with a low denoise of 0.45, generates images that do not change the character, merely changing their eye colours; this is the same when the denoise value in the manual-inpainting group is 0.55. Most of the images you see are generated with the denoise value of 1.00 (the max), and some at 0.75. Context value for most of them is 1, with a few at 1.5 and 1.61 — though I don't think context value matters here, given.... well, the context. The "good" ones you see, and the numbers I mentioned I reverted to previously, are 1.00 manual-inpainting denoise, 1.61 context factor, no controlnet."
- The effect that the denoise has, is very heavily dependent on the sampler & scheduler, they are basically what really drives how much change is introduced after the denoising. When you do a 1.00 denoise basically nothing of the original form of the image is kept, even with a high context area you'll still run into the problem that your inpainted area will not fit into the total image.

"I notice your workflow was done in IMG2IMG, but am I mistaken in assuming that, since it's just the manual-inpainting version, it's still applicable to the TXT2IMG workflow?"
- Yeah no worries, i just had the IMG2IMG open - both versiosn work the same for the inpainting part!

"Moreover, I see that your colour fix strength and hi-res fix denoise are different between the manual-inpainting and the general function control group; is there a rationale for that?"
- The Hi-Res fix in the inpainting area still works the same as before, so lower denoise for hires fixing is required. This is due to it not making a lot of sense to use the Ultimate SD Upscaler since we only inpaint smaller parts and the KSampler is sufficient for that, but yeah maybe i should explain why i recommend different strengths :D

"Lastly, some controlnet-related questions. I realise that you don't have pre-processing turned on — whilst I have read the controlnet tutorial at the bottom, I can't say I really understood what it meant; what difference does it make to turn on each of those options? Same question for the ControlNet group of "Pre-processing image to Pose" and "Load ControlNet Pose". Finally, in the workflow you provided with your settings, since the control net model name is cut off due to length, are you using Depth Midas for both the model and the pre-processor?"
- In the Inpainting Area, if you active "Use ControlNet?" it'll take the inpainted area, preprocess it with your selected preprocessor model and then use controlnet to keep the shape. Any of the other 3 following settings ONLY affect the external ControlNet Image. You can use the external ControlNet Image to upload another Pose, so if you'd mask the arm and then upload an image of an arm waving for example, you could activate "Use ControlNet?" together with "Use External Pose?" to use an external image to influence the controlnet, instead of using the original image. If you don't have an already finished pose for the Image, you could activate "Use Pre-Processing" to instead pre-process the image of an arm, to a pose and then have controlnet use it. Lastly you could also activate "Fit pose into mask?" to move your external controlnet image ONLY into the mask, and not onto the whole image. Otherwise you'd have to fit the aspect ratio of the image and have your wanted pose in the correct spot. With all of these settings activated, you could (for example) use the External ControlNet Image to upload the image of a waving arm, use pre-processing and then fit pose into mask to change the gesture of a person. Like you can see here.
In the General ControlNet-Section of the workflow you can do each individually. You either Upload an actual Image in the "Pre-Process image to pose"-Node if you want to turn an image into a pose and then use that pose to create your image with controlnet (you can see the pose in the ControlNet Pose Preview-Node in the Debug Section) or just upload an already existing Pose in the "Load ControlNet Pose"-Node and use that instead.
And yes, if i want to keep shape and form but no finer details i usually use depth midas for both preprocessor and controlnet.

"Moreover, I've been testing the nmkd upscale model you recommended, and from my experience, it seems like its output makes it relatively blurry in comparison to realESRGAN; or perhaps one could say the latter makes the image much sharper(?). Would you say this is consistent with your experience?"
- Can't say that this matches my experience :/ Here you can see esrgan and the nmkd-siax-200k, while esrgan isn't bad, nmkd for me is usually sharper and delivers better outcomes for me once hires fix was applied :s

@last27683599 No biggie!
Yeah pre-processing is always related to an externally provided image ^^ If you turn on pre-processing without turning "Use External Image" on, it shouldn't do anything and still take the own image :D
Maybe i should update the explanation to make it more clear.

Yeah i have tested anzhc's ultralytics models for a bit and i like a few of them, usually the ones that i release the workflow with are just the ones i have most recently used.
I started dibbling a bit in realism recently and anzhc's detection models were just better than the others since it works for both realism as well as anime quite well.
In a future update i might add them to the recommended list :)

The denoise values that i ship the workflow with are all subject to change, i don't really think much about them, they are usually just the ones that work best for me.
As i explain in the mini-guide there is no real one-fits-all solution when it comes to the denoise value, i usually keep the value at around 0.15 when the generated image already had great facial features and i just want to upscale the quality of that area. When there are some minor deformations i usually increase the value to around 0.35 which is enough in most cases.
Most models now do faces and eyes really well, that's why i use low denoise values that will most likely work out of the box for most people ^^

And don't worry - i never mind answering questions :D ♥

@last27683599 aaaaaaah sorry :D that's what you meant.
Yeah, i usually recommend fiddling with the settings as explained here.

One think i'd always recommend is setting guide_size to 1024 and max_size to 2048, that'll inpaint the area at a higher resolution and resize it afterwards, usually leading to better results.
It's just a bit heavier for GPU's but most consumer GPU's should handle this without issues.

Hey @BiggestFan,

try switching to the newly released v5.1!
ComfyUI has done 2-3 Huge updates in a row that broke even some of their own core nodes, the 5.1 has the same exact functionality as the 5.1 visually, but under the hood it's a lot more stable!

Hope you'll enjoy the workflow ♥️

@BiggestFan happy to hear! 🥰

Tje regional prompting basically already works using the manual inpainting area.
The workflow from Potatcat basically doesn't do anything different there.

Just activate the manual inpaint area, and then activate the controlnet+preprocessor to lock in the pose and position. Do not put the strength too high, somewhere between 0.45 and 0.55 should be fine. Then draw the mask of one character -> write the prompt -> hit run and it'll replace the person using the mask and the controlnet.
Then just hit the refresh button (or leave auto-refresh on) and the new image you created previously will be loaded. Now just mask the other character and write a different prompt and hit run :)

That should work just as good, sadly adding multiple more nodes would make the workflow even more complicated and i'd prefer not to put too much into it, else i'll run the risk of cluttering it to the point nobody (possibly including me) understands it anymore 😂

Heya @BiggestFan,

That means you have probably installed the RES4LYF Clownshark custom node

Please check the second problem here in the faq section:

https://civitai.com/articles/15480#faq:

This is sadly not something i can fix and would have to be fixed by node-creator.

@BiggestFan no problem! :D ♥️

Yeah i do, i mainly use it to give back to the people who create images and videos using my workflows/models 🫡
No need to gift me Buzz though, always happy to help ~

Hey @frank5171628626,

sorry that it takes longer now, i have mentioned this in the changelog.
But i get that if you're doing larger amounts of images it's really inconvenient.

I have swapped out the default ComfyUI Sampler with the Ultimate SD Upscaler, it takes longer but the results are way better, especially at higher denoise values.
Before the denoise had to be kept really really low somewhere between 0.05 and 0.15 and even then the default comfy sampler caused major image shifts compared to the upscaled image.
To keep images closer to their actual pre-hires fix preview i changed the way we do this.
This also makes it's behaviour closer to the one in A1111 Webui.

BUT maybe my settings weren't really optimized for speed and i apologize for that.
I have tinkered a bit further and if you're generating more pictures, would you mind testing something for me and letting me know if the speed has improved again and if the images still look good across the board?

Could you open this node in the workflow by clicking on the bubble to the left and then on the icon in the top right of the node?
And then in the node in the middle can you change the following 2 values:

- mode_type: chess -> Linear
- seam_fix_mode: Half Tile -> None

This has cut down generation time for me in half, without any visual drawbacks.
If you would be able to confirm it - i'd be more than happy to update the workflows here on the site!

happy to hear @frank5171628626 !

Did you notice any quality drawbacks! if not i‘ll ship the change :)

@frank5171628626 thanks for the feedback! I'll implement it and push the changes for everyone ♥️

hey @keima1203721,

please check die FAQ mentioned in the description.

Hey @vhsk2005560,

This usually happens if the models don't do very well with denoise, you can try two things:
- Change the model and see if the problem persists, if that doesn't work try to deactivate color-fix(if activated)
- Lower your denoise, if it's already at 0.10 or 0.05 and the problem persists, it's something different

I have one more thing you can try if neither of these work, but let me see how it goes first 🫡

@vhsk2005560 what do you mean? :o

@vhsk2005560 Aaah you mean the Assets Tab :D

If your images do not appear there, check if your images appear in the output folder. If they are in the Output folder but do not appear in the Assets tab, it‘s most likely due to a comfyui update :/ They have broken a lot recently and i have experienced the same issue on some instances with different workflows.

They really broke a lot and did some fixing that broke other stuff :D

Hey @nonoiru,
no worries - i hope the translator will help you understand everything ^^

If there is no image being generated it's most likely due to you not having filled out some of the required nodes that need a selected model.
The new ComfyUI versions have a red marker around the nodes that need input by you, if you have an older version here is a screenshot of all the nodes you need to fill out.

You 100% have to fill out all nodes on the left side in the "Model Backend"-Section, just select any model for the "Refiner Backend", if you don't use it - it won't matter.

You can ignore the Detection Models as long as you don't use the detailer.
But you will have to select a Model for the IPAdapter and ControlNet area, you can also just use the little arrows if you don't have any models there to set the value to "undefined", that'll also let you run the workflow.

@nonoiru glad it's working now! Hope you'll enjoy the workflow. 🥰

Heya @hakusei666,

if you don't have an image in the output folder, it means that the Image Saver node was never reached due to an error before that.
Could you share what your comfyui console is saying?

@hakusei666 hmmmm that's weird :o
Maybe there were some models you hadn't selected yet or that hadn't defaulted to undefined yet. These are the model's you should select, even if you don't use them, to ensure that comfy's internal execution validator will allow full execution.

Aww that's nice to hear! Thanks a lot 🥰
I try to make the workflow as easy userinterface-wise as possible in comfyui - even if it really has it's limits ^^ Hope you'll enjoy the workflow!

Hey @Captain_Goon,

while this does sound funny - it's obviously not the intended goal :D
In the Debug section can you unfold the Nipples and Vagina node to check where this problem occurs? If it's in the nipple section, it means that the prompt you're using generates vaginas.
If it's in the vagina section, it means that your selected ultralytics model recognizes the nipples as vaginas, in that case i'd recommend switching that ultralytics model out to a different one.

Hey @flyingdargon341,

whats your comfy and frontend version? ComfyUI recently broke a shit ton of stuff with their new frontend deployments 🥲 i guess they do more stuff with AI nowadays and with less oversight 😅

The broken frontend sometimes connects/disconnects existing node connections or shuffles them. Causing wrong inputs into nodes that shouldn‘t be possible to begin with.

I have setup a new instance today with comfy version 0.18.1 and ComfyUI_frontend version 1.42.8 and that works well, maybe consider updating your comfyui to these versions.

You can find your version in comfyui under settings > about.

@flyingdargon341 yey nice!
Glad it's working now 🥰

Hey @Jeny,

you are actually right, it seems that the current way of the workflow does break during hiresfix when a vpred checkpoint is being used.
I guess it's something about the architecture of the vpred models, let me check if i can find a fix really quick :)

@Jeny can you check if these work better?
https://www.mediafire.com/file/et9ryk3c68qr4bd/TXT2IMG-ADetailer-v5.2-vslinx.zip/file

For V-Pred you might have to reduce the hires fix denoise a bit ^^

@Jeny yay! thanks for testing ^-^

Hey @tal0n,

ComfyUI has recently pushed a lot of vram optimization stuff where it'll offload things to the RAM when it thinks you do not have enough vram available.
There are some instances in the past where i also had this issue, what usually helps is to check that you do not have much vram usage active.
Stuff like Wallpaper Engine for example also do use quite a bit of vram, you might want to rightclick and pause the animation if you're using it.
Other things that could help is to update your comfy backend and frontend to the newest version and/or deactivating the memory optimization by appending this start parameter to you comfy start:
--disable-smart-memory 

@tal0n that‘s weird and really shouldn‘t happen on a 4080 :/ that sucks

In the TXT2IMG workflow, in the bottom left corner inside the Backend Section you should she a „Sampling“ Subgraph, if you open it - you‘ll find a green group with the normal Sampler. Here you should be able to replace the VAE Decode to a tiled one.

In the IMG2IMG workflow you will find the KSampler + VAE Decode in the IMG2IMG Transfer Subgraph at the bottom of the workflow inside the „Image transer“-Group.

Hey @dNerop,

glad you're still enjoying the workflow 🥰
Yeah comfy can be a bit less user-friendly than a1111, forge or swarm ui ^^ which is why i'm trying to make the workflows as feature-rich but visually easy for everyone.

You actually can use the wildcard selector to add multiple, just click it again and select a different wildcard, the selection down there will just show you the last selected one.
But i understand that it might not be absolutely clear UI-wise that you can add one than more, so i might make it so that the selection jumps back to "Select a Wildcard" after selecting one.

@dNerop thank you for recommending my work to others! 🥰
I have pushed the changes to the node 1-2 Days ago, so if you update the vslinx-nodes through comfy manager you should now have it so that the input field resets to "Select Wildcard", which is the more user friendly approach :D

Heya @Feral_Signal_Studios,

thanks for the compliments first of all!
And no worries - i have answered the question a few times by now but i don't expect people to go through old comments or anything ^^

I have in the past not intended to implement IPAdapter for Facial identity.
I think it's better to leave that to tech like faceid/instantid etc. and to have that as part of it's own dedictated worklfow since it's hard to implement into the workflow in a meaningful way that wouldn't overload the UI even further.

But i'm currently working on the 5.2 version of the workflow and i have removed the ipadapter for style and composition and instead i'll leave it to the user which ipadapter type they'd like to use. So there will only be 1 image input and you as a user can decide which ipadapter type you'd like to use.
So you could use ipadapter for facial features as long as you pick the correct type and then download+select a ipadapter model that is specialized in that regard.
You can always find my work-in-progress versions of the workflow on my huggingface repo here:
https://huggingface.co/vslinx/ComfyUIDetailerWorkflow-vslinx/tree/main/workflows/TXT2IMG/v5.2

You will have to make sure you have my vslinx-nodes completely up-to-date though, i have released v1.9.0 today, but comfy registry and comfy manager sometimes take up to a day to actually pull the new version, so you might have to directly pull the newest version of the node from the github of my node here:
https://github.com/vslinx/ComfyUI-vslinx-nodes

@hakusei666 i'd start trying by just using linear at the beginning if nothing has been recommended on the modelpage itself.

Usually when applying ipadapter for likeness of a character you'd want to use either Linear or strong middle/ease in-out if you want stronger identity lock-in without it bleeding into overall composition/style.

Hey @Rocco87 and @insaneriverg732,

this is not an issue of the workflow - but sadly a problem caused by a node update from the image filter :/
The node-developer has pushed a change that made the node downwards incompatible.

You can fix that easily though by rightclicking the "Image Filter"-Node at the bottom in the Sampling group. And then clicking on Fix node (recreate). Like you can see here: https://qu.ax/x/lSbuJ.png

@Rocco87 hehe, glad to hear it's your favorite workflow 🥰
I'll push v5.2 soon that'll come with it pre-done and some more changes.
This already happened with the image-filter once in the past, so i might just create my own version of it with backwards compatibility 😅

Hey @Vaisoul,

TXT2IMG means you're turning your text into an image, so you create a new image from scratch.
This workflow has built-in processes to directly refine said image based on your settings, or not at all if you don't wish to.

The IMG2IMG variant is for either refining images or creating new images based on an already existing image.