š ļø Purpose & Design Philosophy
This workflow is a high-fidelity environment built for Illustrious. It prioritizes stability and professional texture over generation speed. It follows an "all-in-one" philosophy: configure your prompts, hit queue, and let the workflow handle the multi-stage refinement from start to finish.
Not for Speed: This is a heavy-duty refinement tool. If you want 2-second previews, use a basic workflow.
Personal Use: Built for my specific production needs. It is shared as-is for those who want a "set-and-forget" pipeline for Illustrious. Adjust the settings and models to fit your needs. The defaults on v19+ will be what I used for the sample images.
All-in-One Logic: The workflow handles generation, detailing, and upscaling in one continuous pass.
If you're needing something with more features: Try checking out Silly All-in-one, Multiple Characters With Regional Loras which is on civitai. I have no plans to implement additional features.
ā ļø Disclaimer & Compatibility
Install at Your Own Risk: Custom nodes can break your environment. I am not responsible for troubleshooting your specific installation.
ComfyUI: Built and tested on the non-app version. Desktop app users may face additional hurdles.
The "Your Version" Factor: Your node versions and environment are 99.9% likely to differ from mine.
Nodes 2.0: I do not recommend using Nodes 2.0. It creates unpredictable UI behavior; I will not provide support for issues involving this feature.
š¤ Support & Boundaries
I have not and never will monetize my workflow. CivitAI is the only platform I am maintaining this workflow on. If you find a version of my workflow posted and/or monetized anywhere and use it, then whoever posted it is responsible for providing you any support. This is the main reason why I don't monetize this: I don't want to be or feel obligated to provide support.
No DMs: DMs are disabled due to repeat spam. Please check the Discussions tab below; most questions have already been answered.
Modifications: You are free to hack this workflow apart. However, you are responsible for fixing it if it breaks.
Custom Requests: I do not make private workflows. If you need a custom solution, post a Bounty on CivitAI. There are many talented creators ready to help you for a fee.
Description
Version 18.0: The "Streamline" Update
This version marks a major architectural shift. Taking lessons from my separate Anima project, I have gutted the "bloat" to focus on a cleaner, faster, and more stable core experience.
Key Changes:
Model Patch Overhaul: Removed Epsilon Scaling and SMC-CFG. These have been replaced by a single RescaleCFG node. Combined with the existing CFGZeroStar, this provides superior "Anti-Burn" protection and prompt adherence with much less complexity.
Wildcard Removal: Removed the ImpactWildcardProcessor. As part of the effort to simplify the prompt path and reduce node dependencies, the workflow now utilizes standard CLIP encoding for more direct control.
Refined Upscale Logic: Removed Whiterabbit nodes. The pre-USDU upscale group now utilizes ComfyUI Core nodes. This offers a cleaner path with a simple toggle to choose between an Upscale Model or Pixel-only (Lanczos/Bilinear) upscaling.
USDU Optimization: Removed the Noise Injection group from the Ultimate SD Upscale process. Testing showed the benefits did not justify the extra complexity and clutter.
Workflow Cleanup: Removed Image-to-Image and ControlNet modules. To keep the workflow focused and readable, Iāve stripped these out for now. They may return in a future "Expansion" module, but the current goal is a lean, mean generation machine.
Watermark Handling: Removed the Watermark Detection & Removal group. Improved negative prompting (using "username") has proven effective enough to handle this natively without extra nodes.
Scheduler Simplified: Removed the Laplace Scheduler. While interesting, the extra wiring required to make it toggleable and compatible with metadata saving was causing unnecessary workflow "noise."
Generation time on a 5060ti 16GB: 2 LoRAs loaded. 140 seconds from start to finish. Image Gen>Upscale>USDU>2 Detailers.
FAQ
Comments (19)
v18 looks great so far. The workflow is much easier to read without the excess of advanced options and the extra tips are helpful. One point of feedback from my initial tests is that CFG Zero Star + Rescale CFG work like magic with some models but can be a little much with others. It might be safer to keep a bypass on those. The Detail Daemon can technically be bypassed by setting detail_amount to 0 but it doesn't hurt to keep the bypass. So far it does seem pretty safe to just leave on with the default settings.
The Impact Pack supposedly has a CustomSamplerDetailerHookProvider node which lets you connect a custom sampler to the Detailers. Strangely, this didn't seem to have any effect when hooked up to the Detail Daemon on my end. Figured I'd mention it just in case.
Yeah, I tried the before with the detailers. Didn't seem worth it to me.
Regarding bypasses It's less clutter just to have folks use ctrl-b or click on the node and click the bypass icon. I'd have to put those nodes in individual groups to make them easy to bypass, which then eats up more canvas space. The main goal with v18 was decluttering. Folks can add bypasses if they want them if they really need them right now.
I might do it in the next update since I already added Laplace to my personal version of the workflow.
Today is a great day, thanks G!
Your workflows do wonders for my learning phase.
Do you happen to have an older workflow with inpaint?
You are doing an awesome work in your workflows, the v13 was very good, but the v17 was an huge step, the ultimate sd scaler is God and the controlnet works so good. I havent tested the newer ones.
Thanks!
v18 doesn't have ControlNet or Image 2 Image, so it might not be up your alley.
laplace v4 is working woonders for me so it is worth trying the v18 or just keep laplace since it is workinmg really well?
v18a is what I am using now.
Is it worth trying? Sure, but if you're happy with what you're using already, then there's no need to switch IMO.
@zerol1 For having tested many versions, if you like tweaking and experimenting, go ahead you'll have fun!
If speed is important, I've seen some good improvements in some recent versions compared to the old ones.
@GladasĀ @FellaitioĀ thanks a lot =D i will try v18 i hope it dont break my install lol
@zerol1Ā if that can reassure you, I had no issues using any of the versions. Just a couple of nodes might be missing and not automatically installed by ComfyUI Manager which, in this case, requires a bit of search. At worst, I would recommend doing a 2nd portable installation to avoid messing with your OG one.
First I wanna say thank for these workflows, absolute genius you really made something magical. Now something I noticed when using your workflows, everything worked great the first time. Then after repeated attempts suddenly when the upscale is finished, the image looks messed up with faces on body and arms somewhere else. I haven't changed any setting and this is prevalent on all workflows you've made. Im not sure what the problem is, i tired asking good ole AI for help, something about the tiling but despite changing stuff around it still produces these errors. Such as extra limbs, hands in other places, even triple breasts and double nips, My negative already includes this stuff as well to prevent it from happening but nothing.
Take what I say with a grain of salt since I am basing it on the information supplied in your message. Also, I am not ComfyUI dad, so I can be 100% wrong:
I want to start off by saying, that I built and tested the majority of my workflows for 2x upscaling. If you're going higher than that, you will probably run into heavy hallucinations if your tile_padding isn't high enough. Based on some of the stuff you are saying, this is probably the case.
Here's a quick note from Gemini about tile_padding:
"Padding allows each tile to 'see' a bit of the surrounding area. Without enough padding, a tile containing a sleeve doesn't know there is an arm inside it, so it might turn the fabric into a new face."
I usually use 256 tile_padding during my 2x upscales to play it safe, but I will do 128 if I want it to be faster and am feeling lucky.
If everything was working fine and the only thing you changed was your prompts and not the upscale settings, then the issue points to your prompts. Everything you prompt (positive and negative) is potentially giving USDU more things to hallucinate. Avoid repetitive and redundant prompting. This is not me attacking your prompts, this is just something I have noticed in my own experience with USDU.
You can try giving USDU it's own positive/negative prompts by adding CLIP Text Encode nodes and connecting them appropriately. Some people say this helps, but in my experience it sucks. The USDU only prompts should have minimal prompts. (e.g. just the quality prompts).
Some things can't be saved in USDU no matter what you change as far as settings go. When this happens, just give up. Otherwise, try some other workflow/method of upscaling if you are dead set on saving that image.
Plenty of people ask me why I use USDU instead of a 2nd KSampler or <insert other upscale method here>. 2nd KSampler (aka hi-res fix) is more stable, but not as good on adding detail (IMO). Other people will argue the opposite, but as I have stated, the workflow is designed for my use and is not guaranteed to work for everyone else.
I have also tried things like SUPIR and several other upscale methods. They are either too complicated, too resource heavy, or some other random reason why they sucked (IMO).
The simplest solutions if you don't want to try adjusting prompts and/or other settings:
1. Try a different workflow. Maybe mine just isn't a good fit for you.
2. You can try swapping out USDU for the Hi-res fix method via KSampler (KSampler, SamplerCustom, SamplerCustomAdvanced, etc.).
Someone else may chime in with an "ackshully" comment, but as I have stated "I am not ComfyUI dad."
Google Translate
When I used version 13, I had a similar problem. Try fiddling with the settings tile_padding, seam_fix_padding and try using Embedding
Epsilon Scaling will likely be added in the next update.
I have been using it on just the initial image and it is giving me the lighting and contrast that I like without having to lean on Laplace or using LoRAs.
A brief note explaining the scaling_factor setting will be included on the workflow based on the ComfyUI release notes and revised by Gemini.
If you want to try it yourself, here are the notes:
Epsilon Scaling (Exposure & Detail Correction)
Unlike standard scaling nodes, this node is used to attenuate (reduce) the model's noise prediction to correct for "Exposure Bias." This brings the image closer to the model's "ideal" mathematical state, resulting in finer textures and better prompt adherence.
Key Settings (Highly Sensitive):
scaling_factor: 1.0 (Default): No effect.
* 1.001 ā 1.020 (Recommended): The "Sweet Spot." This slightly reduces noise prediction, which sharpens fine details like lace, satin, and skin pores. Even a change of 0.005 is visually noticeable.
* 1.03+: Can lead to a "soft" or "under-denoised" look. If the image looks blurry, your scaling factor is too high.
Placement:
Must be placed in the model patching chain (between *Load Checkpoint** and your KSampler). I place it right before the other model patches.
Why this works for my workflow:
Because you are using RescaleCFG at 0.7, you are already "taming" the brightness. Adding Epsilon Scaling at a low value (e.g., 1.005) provides the "micro-contrast" needed for high-fidelity textures without the speed penalty of PAG.
Note: If the image feels too "dark" or "flat," lower this value back toward 1.0. If it feels too "noisy" or "jittery," increase it slightly toward 1.01.
I would like to ask if you have thought about creating a regional prompt?
@Wanderer_000Ā probably not. I have never needed it and I feel like it would over-complicate the workflow. Epsilon Scaling is a single node that people can just delete or toggle off.