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
FAQ
Comments (20)
i really like the workflow already but i think im having an issue with a part of the detailer groups.
each one of the SAMLoader (Impact) nodes highlight red when i run the workflow, and i'm unable to choose a model_name.
everything else seems to be working good. any help is appreciated, thx.
@interstellar177869 you either have to download detection models or remove the detailer groups from the workflow if you don't plan on using them. There are tons of detection models on civitai or you can download them via ComfyUI manager under model manager and filter for ultralytics. If it's the SAM model, then you need to download and install one. Easiest way it to use comfyui manager/model manager - filter for SAMS
@GladasĀ thx, i figured out the issue; the folder i was putting my models in was just named "sam" while it needs to be "sams." this probably could've been avoided if i used model manager so learned my lesson there.
@interstellar177869Ā glad you got it figured out!
Hello. Based on comments from people who have used my workflows, most seem to like v13. If you're asking what I use: Laplace v4 (as of this message).
Otherwise, you can look through the various versions and decide what sample images look the best from your point of view and then try them yourself.
As far as upscale settings go: you can set the upscale to whatever you want as long as your hardware can handle it. Just be aware that you may need to set the tile size manually. For the most part, my workflows have the tile size connected to the original image size, but the newer ones are tied to the size of the image going into USDU.
USDU likes to hallucinate depending on your settings and/or prompts based on my experience. My images are 2x upscaled and sometimes I still run into USDU hallucinating.
@GladasĀ Thank you, I'm using a modified version v11. I think I added two more upscales after the first one and between the second, and it reduced the response time of the detailers, as well as a mini-system to generate images based on the other. The newer versions are much faster, mainly because of the new LoRa system (LoRa Loader), but for me there isn't that much room for customization for more upscales, etc. That's why I asked if there was a balanced version between them. I'm going to test v13, Best regards.
@DexsCivĀ The easiest to modify will probably be v15 since nothing is pinned, but "easy" to me doesn't mean "easy" to other folks most likely. In any case, hope v13 ends up working for you!
@GladasĀ I was testing and wondering if it's possible to go from one upscale to another and so on before going to the USDU? Currently I use a kind of split load between 2 upscales and 2 USDUs, which creates an excellent 3D image but also some pretty hallucinations.
@DexsCivĀ by "upscale", I assume you mean where it's upscaled via upscale model. It's possible to do. If I remember correctly, I no longer use upscale models in my workflows as of v15.
Short answer is yes, feeding an image through multiple upscales before USDU can be done.
@GladasĀ Any optimal settings that I should focus on to avoid hallucinations? I've made gems with your workflow, so thank you very much, but I to tone down the upscale to 1.5, denoise is now super low and yet, it does hallucinate almost half the time. Any tips and tricks would be very appreciated!
@FellaitioĀ unfortunately, there is no magic formula to avoid hallucinations with USDU. You're basically experiencing USDU being USDU.
I am not the authority on USDU, so what works for me doesn't necessarily mean it will work for you.
What I do:
Some tags/prompts are stronger than others. Avoid describing the same stuff in different ways when prompting. Excessive background prompts will definitely cause USDU to hallucinate. Avoid using tags like 1girl and 1boy when possible. (Alternatively, you could add and connect separate CLIP Text Encode nodes to USDU and use minimal prompts on those , but this has never worked for me personally.)
Adjust your CFG setting, try raising or lowering it. (Usually lowering it)
Adjust the amount of steps. Try using less.
Try a different sampler or scheduler. Certain samplers and/or schedulers are more prone to hallucinations.
Try a different model/checkpoint to compare.
Avoid excessive loras.
Bump the tile_padding up: 128, 192, 256, 384, 512, etc.
Try hitting up your favorite AI chatbot and have it troubleshoot with you. They're not always right/wrong, so take what it says with a grain of salt.
Other tips:
If you're manually tweaking the tile width/height, make sure they are a size that is supported by the model you are using.
You can add ControlNet to USDU, but expect longer gen times. I used this back in the day, but not worth it to me.
@GladasĀ Man wow! Even if it might not help much, thanks so much for writing all of this! You did include about half the things I already tried but you added way more stuff I can't wait to experiment with! Thanks so much! Want me to report back with what did improved the most or not needed?
@FellaitioĀ I'm all for sharing info, since it might help someone else in the future, but please do not feel like you need to.
@GladasĀ @Fellaitio During detail upscaling stages, it's advisable to configure scheduler for less variability in low-frequency noise; in the simplest case, use Exponential with a noise level of around 0.18.
Sometimes the issue isn't hallucinations, but rather incorrectly set parameters that over-process low-frequency noise.
Higher sigma numbers will result in more variety in the large objects, while smaller numbers would mean less variety with finer details.
Pay attention to the frequency characteristics of the processed noise. You can use SigmasPreview (RES4LYF) to visualize the tensor graph.
Any chance we get a version without the Comfyroll (CR) nodes? Looking at the github page it seems like the project is abandoned with 100+ issues raised and multiple unmerged PRs needing to be manually applied. Seems like a horror show to maintain compatibility into the future...
@MeitanteiJesus If they stop working, then I will. So far, I have had no issues with them in my workflow. If you have suggestions for a single custom node pack that does all the stuff that the Comfyroll nodes do in my workflow, then feel free to mention it here and I will take a look.
For me: if it ain't broke, don't fix it.
I'll do some digging on my own, but if it becomes too much of a hassle for me, then I just gonna stick with what I have.
This turned out to be a HUGE pain. Comfyroll was basically the backbone for many parts of my workflow. I will only be updating on the latest workflow "series" (Laplace) since that is the workflow I use.
@GladasĀ Thanks a lot! I was actually in the middle of working with LLM Agents to fix all the broken bugs in Comfyroll, but it definitely felt like every time a python library gets upgraded, some compatibility issues need to be fixed. I'll try to study the workflow now and come back if I have any questions ChatGPT/Gemeni/Discord can't answer :)