This is the start of my 'new' workflow(s). We'll see how long they last before they get nuked by the powers that be.
v1 does not support AIO models by default. (Models with VAE and/or CLIP baked-in).
Your model (not VAE/CLIP) need to be placed in your models/diffusion_models folder to work with the model loader in the v1 workflow.
This workflow has the following functions/features:
Text to Image
Multiple Lora Loading
Global control for Seed
Wildcards
Upscale
Detailers
Color Correction
Save Image with full resource metadata in CivitAI readable format
This workflow is not setup to work with models that use CFG higher than 1.
Support Policy
These workflows are provided as-is and without guaranteed support.
I do not provide support for third-party applications, modified versions, redistributed copies, reposts, bundles, or installations obtained from unofficial sources.
Inclusion of these workflows in another project does not imply my endorsement, involvement, or responsibility for that project.
Disclaimer
Responsibility for generated content falls on the user of these workflows.
Installation of ComfyUI and/or any related custom nodes for these workflows are at the installer's own risk.
Description
Initial release.
Some minor changes from my old workflow (if you were using it).
Removed CFG Zero Star
This does nothing at CFG 1.
Removed the text box for the Negative Prompt
Negative prompts are useless at CFG 1, so this is just decluttering.
Image Comparers now show the 'after' result by default.
FAQ
Comments (15)
Awesome!
Let's go! Welcome back. 馃榿
If you're getting results that look like there's severe grain or static, try deleting the Conditioning Zero Out node connected to the red CLIP Text Encode (Prompt) node. The image will still generate without it.
Simple and effective workflow, as an avid wildcards user, this is exactly what I was looking for. Thanks!
Pretty much every image I've ever posted has come from a Gladas workflow! 馃挋
I've resorted to remove using USDU and just went with a much higher resolution from the get go straight from Krea2. I get similar results but without any skin cracks/extra skin wetness and no hallucinations.
Mentionning it in case you never tried and wanted to try on your end to see if it does output better results
I have tried it, but it's kind of 50/50 for me.
Doing a bigger initial image is going to always be better IMO, but not practical for everyone.
You can technically use USDU as a sort of second KSampler by not upscaling. Let say for example you generate your initial image at 2048x2048 and set the upscale to 1.0, it will just act like you're doing IMG2IMG essentially with whatever denoise setting. The "tile" will be the same size as the initial image, so it's just one single tile.
My personal experience is that many of the Krea 2 finetune models just don't perform well with USDU in general. Project Chimera has been pretty decent for me in USDU, but other ones are not as consistent.
Another alternative is to have a 2nd KSampler instead, but this has given me not so great results, though other people's experiences may vary.
@Gladas聽I get it that it might be better for some folks but I also reduced my time to generate to roughly 40-50s for a 1664x2432 image.
If you find a way to mitigate completely the skin cracks and excess texture (especially on skin) from USDU, I'd love to know how ^^
Something i posted if you want to take a look and see people's comments:
https://www.reddit.com/r/StableDiffusion/comments/1v5oayu/krea_2_usdu_skin_cracks_weird_textures/
@Fellaitio聽unfortunately, in my experience there is no universal set of perfect settings for USDU. The behavior varies by many factors including the model. I'm not trying to convince you that USDU is better, because it is not for everyone. The overall generation time will definitely be faster when not using USDU, regardless of hardware. I can take look at the comments on your link, but I already assume people will be like use SeedVR2, use a 2nd KSampler, or 'Krea 2 doesn't need upscaling'.
@Gladas聽It was more to show you different suggestions/techniques but yeah, those were also mentionned! I really wanted to keep USDU tho, for authenticity with your amazing WFs but for Krea2, it seems I haven't been able to make it work properly for semi-realism
Seems the replies were actually meaningful in that thread with the recommendations of lower steps, etc. Definitely worth trying with USDU (if being used).
I sincerely hope people are adjusting settings to their own preferences instead of thinking that the workflow is set to make a perfect image.
One person does mention something similar to what I said about using USDU with a single tile.
@Gladas聽Just wanted to share if you were experimenting new ways for your next version (if there's one coming, that is)! ^^
@Fellaitio聽the great thing about workflows is that they can be modified. The thread was actually pretty good and not full of troll replies. I won't be adding SeedVR2 though. People can do that on their own lol. I think the single tile approach is probably a solid approach, but still not feasible for all.
@Fellaitio just wanted to follow up with you. I had ChatGPT try looking up any new guidance on USDU with Krea 2 Turbo and also had it look at this conversation plus that reddit thread you linked.
TL;DR - it recommended using square tiles instead. For me it recommended using 1536x1536 tile width and height. Since I typically only am using SDXL resolutions, I can actually drop to like 96 tile padding without any issues, but it recommended using 192 tile padding for more general use.
I ran my testing using:
1. Initial Image 832 x 1216
2. 2x upscale with no upscale model.
3. Euler + Simple, 4 steps
4. 1536 x 1536 tile width/height
5. 96 tile padding
The results were clean for me, but YMMV.
I think that if the method you are using now works great for you, then stick with that method.
This method takes about 120 seconds on a 5060ti 16GB.
@Gladas聽Oh damn! will def try when I get the chance! Thanks so much for the detailed info, I'll report back just for fun (fully aware that my mileage can vary so no expectations! Will be sharing for the fun of it!)