No styles, no detailers, raw and fixed.
This is base NoobAI-XL v-pred 1.0 with my colorfixes merged in at a specific weights.
This is v-pred SDXL model. Read guides on what UIs support it and how to use it.
Making it was a tug of war between desaturation and oversaturation. Sliders trainer modified to run v-pred, all versions of torch and FA from 2.3 installed, 70+ loras baked, constant back and forth and stuff not working as I thought it would.
But now everyone can see what I found in this particular model and why I spent all this time fixing it.
This model inherits original's license.
This description will probably be updated later.
Description
I felt colors were too washed out, so adjusted contrast a bit. Now blacks are more black, greens are more green. Dark is a bit darker too. It generally feels even more stable now somehow.
FAQ
Comments (15)
Hi! Thank you for sharing the new checkpoint.
Just to confirm — if I use this checkpoint, does that mean I don't need to apply the NOOB-AI XL V-PRED STOOPID COLORFIX toolset anymore?
I'm assuming the toolset is already merged into this checkpoint, but I wanted to make sure. Thanks again!
Yes it is, so there is jo need. But you can apply it to make look overfitted again 😁 Also you csn apply it to offset loras that produce sam effect.
@Volnovik
Thank you for the reply!
Some LoRAs tend to distort the colors, so I’ll consider using the toolset again in those cases.
I really appreciate you sharing such an amazing model—thank you!
I don't know the reason, it's worse than og model even with RescaleCFG 0.7
I can only look at your workflow as text right now, but it looks kinda bad. You should at least change this: S&R":"ModelSamplingDiscrete","widget_ue_connectable":{"sampling":true,"zsnr":true}},"widgets_values":["v_prediction",false
And set v_prediction to true. Try my workflows as a start
@Volnovik thanks, solved
Thank you for your work, it's truly impressive! I’d like to ask for your advice.
I often create images in a specific style, mostly using flat colors, simple lighting, minimal realism. For that, I’ve been using WAI-shuffle. While the final results are decent, I still feel like the model adds too much realism, with strange shadows and unnecessary lighting.
I’ve spent quite a while tweaking prompts, negatives, and using extra tools like your colorfix slider to try to get rid of those unwanted elements, but I still haven’t been able to fully achieve the style I’m aiming for.
That’s why I’d be very grateful if you could recommend a model that would better suit my needs — I believe that might be the only way to truly solve this issue
Use model that supports flat color or 2d as default style. I'd suggest using this model with specific artist tags, avoid quality tags, add stuff like thick outlines and flat color into positive. Or even minimalism. Following example is just 30% of steps with artist tags, but it is way more flat. https://civitai.com/images/76863139
Also I'd avoid detailers, they tend to add those halftones and shadows.
@Volnovik Do you happen to have any specific models in mind? Preferably something V-pred based, most of my LoRAs were trained on the base NoobAI.
Right now, the structure of my prompts looks like this — https://imgur.com/a/uR5m9GH, I’ve kept what I believe are the most important quality tags. Could those really be negatively affecting the overall style?
As for detailers, I only use yours. Do you think I should stop using it?
@degurshaft Models are just base either finetuned over curated specific dataset of merge of certain style lora or even merge between different models. Illustrious and noob are trained over danbooru dataset with thousands of style tags already baked in. Some of them turned out better, some worse. I did not like the way community "fixed" vpred, basically shifting it towards eps, eliminating most of v-pred capabilities. That's the reason I decided to fix base model itself. The set of my colorfix models is built specifically to battle oversaturation and weird effects that occur during v-pred training. I have no idea what style do you want. You mentioned too much realism, that's why I assumed you wanted a more flat color experience. If you just want style - try chasing for style tags. Here is a quick comparison: https://civitai.com/images/83686043
And here is the qucik comparison of different style tags: https://civitai.com/images/83688105
You can mix and match those artist tags to build a dataset of your own style and train on it on this model, then use my toolset to offset colorification back to your desired values.
@Volnovik Well, I see what you mean. I’ve actually been experimenting with different artist tags for quite a while now, so I guess I’ll just keep at it.
I also agree with you about the general trend in V-pred models. It really does feel like all those soulless merges of Noobai and Illustrious are everywhere these days, and they’re basically the same eps models :(
I reading all of your article and a comments in reddit.
How is this checkpoint better than Janku vpred and waishuffle vpred?
I mostly use epred model and I plan to experiment on three vpred tomorrow.
Sorry for wall of text, but it is what it is.
This is just base model with colorfixes. You can see datasets, you can say it was tuned on bunch of sliders. As I said in reddit post - vpred sdxl tend to oversaturate on... anything. Too much LR, too long training, too saturated images, improper captioning. Just think about it, training text slider leads to this oversaturation. This is either bug in diffusers sampler calculations or CLIP "feature".
This effect looks cool at first but gets old really fast, so this is just base model that is colorfixed. As base illu it has it's issues, training on raw 12M danbooru images has this effect. But it's diversity makes it better as base for training and applying loras. Prior to this consensus was that you either lower cfg or increase rescalecfg to "fix" certain prompts on per prompt basis. That sucked, so I was fascinated how good and bad at the same time noob vpred could be. This is the result of me trying to fix it. My article is more of a guideline explaining HOW to approach fixing YOUR finetune. There are couple finetunes out there, but they are either shifted too much towards eps or, like Rouwei training just made it forget a lot of stuff.
TLDR - it is just base and way forward for other creators to use. It was never planned to be better. It is just few tricks from color theory that really worked. But I exclusively use it now and am amazed by result more than anything else I tried. Though I have certain bias here. Probably will have to make a big comparison to show the difference. In the meantime you can check it out yourself.
Anyway I'll release update today, because I totally missed greens and eliminating them without making image purple or frying was hard.
You can also check https://civitai.com/models/99619?modelVersionId=1965505 this vpred fix. Author is ignoring me, and I am pretty sure he used similar approach resulting in realistic bias. At least I am confident he was muted colors slider at some point, lol. My thing evades that, and I actually had success combining them both. Today I'll release update to this model, hopefully eliminating last muted colors + excessive greens and move to a different merge I was cooking in the meantime.
Volnovik since you know a lot about this stuff.
1.What should I use between your details lora and resharpen extension?
2.I saw you mention about Vram option for controlnet. Do I need to set anything if I use 24gb GPU with forge?
3.What about hires fix? I like to use the hires fix with 0.25 denoising and model such as Remacri or Anime6b. I feel like the hires fix add more detail into the image without destroying the style of the checkpoint.
4.What about adetailer?
Personally I don't like to inpaint, I like to using an adetailer to fix the hands but it broken when the hand holding an object.
5.Out of three model the colorfix is only one that can produce the color that optimal for sRGB monitor but it also tend to output a weird interaction and bad hands. Wai-Shiffle is better for interaction concept and hands/fingers on object with a cost of mid color. Janku vpred1 is mid-low on everything and feel like a style lora better use Janku E-pred at that point.
Is it possible to use your model for a "generate > hires > done" or "generate > hires > adetailer > done" the vpred is giving a low success rate without inpainting compared to epred and for some reason I can't use the same seed and get the same result anymore. must ne DDPM or SGMUniform. So I can't do "generate > cherry pick > re-hires"
roseodimm 1. You can use both.
2. 24gb is enough for everything for SDXL at full precision.
3. Forget about hires fix. My current flow is: generate, fix and add details via inpaint masked, upscale x2 using mod at high cfg + tile controlnet, then second pass just mod at 0.22-0.24 denoise. Preview image for this model was made this way. This way took me around 25 minutes, 15 of which was figuring out something cool and struggling with kontext to add text.
4. Learn inpainting. It is faster then cherrypicking and you can do literally anything. Any pose, any details. Read my guides, they have vital inpainting tips. You can fix anything, sloppy hair, eyes, buttons on clothes etc.
5. Elaborate more. I don't understand. Especially about srgb. You can play with artist tags, but i suggest using at least one good style lora as base for consistency at low weight. This was true for pony and all other models, base is just too inconsistent. Anyways I cannot elaborate further since you do not share your results, maybe it is just prompting issue or you forgot to set rescalecfg.
Waishuffle is just too eps, it is unable to do really dark or colourful imagery. And cannot keep up with details. But it is a subjective opinion, just use whatever suits you. My way of doing stuff is not winning any awards or thousands of likes, so take it with grain of salt