WIP - DO NOT USE
Still in the middle of getting this description and such fleshed out, so WIP.
🐍 bigASP 3
The next evolution of bigASP has begun. A general purpose diffusion model built on the Flux.2 Klein architecture. Photoreal, anime, cartoons, furry, product photos, you name it. As usual, strong performance both on everyday safe prompts and 🌶️ prompts. A model meant for you to express your creativity.
Start with the RL2 version, and follow Usage below for the best experience.
Usage
Use the RL2 model
20 steps
Euler or DPM2
1.0 Guidance (i.e. disabled)
Detailed Prompts
Any Preset Resolution
Drop this image into ComfyUI for a ready-to-go workflow:
Status
bigASP 3 is currently still UNDER DEVELOPMENT. This is a very early ALPHA release. Just to give a taste of the model. The current version has tons of problems and will likely not be useful for your everyday gens.
Pre-training is complete, establishing the base model. Post-training is on-going. The post trained model is stable and usable, but has some odd artifacts (color tint and noise, mostly) and needs a lot more polish and improved photorealism.
Versions
RL2 - Use this to try bigASP 3 out. Stable and usable.
Base - This is the pretrained model with no SFT or RL applied. More creative, far less stable.
Known Issues
Noise/grain - The model loves adding a bit of noise/grain to the images. This is likely the model hacking the "detail" reward during its training.
Green tint - The model is obsessed with the color green, preferring it over other colors and adding it as a tint to many gens. No idea why, but I'll tweak the Color reward for future runs.
Usage (More details)
The RL2 model was trained on 20 steps, DPM2, with no CFG, and these resolutions:
Square: 1024 × 1024
Portrait: 736 × 1472, 768 × 1376, 832 × 1248, 864 × 1152
Landscape: 1472 × 736, 1376 × 768, 1248 × 832, 1152 × 864
So, generally speaking it will work best within those settings when you use it. It should be able to handle things outside of it. Maybe even 2K resolutions. But it will be generalizing in those settings so performance might be worse.
Both Base and RL2 were trained exclusively on detailed prompts, so the model works best when it has most of the desired image specified. I've trialed it with different kinds of prompting, under prompting, etc, and it doesn't seem to have too much trouble when the prompts are in different wording and formatting, but with underspecified prompts it can be less reliable and image quality will suffer.
Yeah, I know, it's a pain in the butt. But first and foremost I wanted bigASP 3 to follow prompts reliably and well. Prompts can be fixed, the image model less so :P I plan to train a prompt enhancer tuned specifically for bigASP 3. So the recommended flow will be: ask for what you want -> prompt enhancer creatively fleshes that out -> bigASP 3 generates it to the T.
For now I recommend using your LLM of choice to help with prompting.
Description
Second RL Test
FAQ
Comments (28)
Can you recommend how not to use it better?
I'll get more details up soon. Use the RL2 version as a flux 2 klein 9b model in ComfyUI, 20 steps, Euler or DPM2, no CFG (guidance set to 1.0). 1MP resolutions. Detailed prompts.
Starting workflow is up in the model description now.
Like a couple of hours ago I was looking back at some BigASP gens I made and wondered if/when a new version was coming. And now here it is. Very cool!
amazing. I was playing with the rough release from HF a month or two ago and it was already incredible, even with issues. Excited to try a stable version.
I am getting mixed results but some are looking really good! Could be my future new favorite model :-)
I take it back... its amazing!
My old new number 1 is back!
Great news! But oh man, I was hoping Bigasp 3 would be for Krea2!
I think the same dataset would be worse on Krea 2 TBH. The VAE isn't as good. And it's a much "stiffer" model to begin with, like it can't do facial expressions properly at all out of the box on its own even.
@ZootAllures9111 yes but it is faster and lighter
@faketitslovr Huh? Krea 2 is not faster than Klein 9B
I thought he would have been working on Klein before KREA2 was open sourced so that may be why, I'd like to see if for KREA2 but clearly it's taking considerable time for Klein given this is a preliminary model so could be a long time before KREA2 but I think Klein is a good model and knowing how radically different BigASP was on the SDXL format I have a feeling he might have trained out the bad things of Klein namely the bad anatomy but we'll see.
Its quite excellent, but at CFG1 it causes at least in my case horizontal lines. At slightly higher CFG, with some anti-overcook it produces just nice image. Unfortunately thats really quite slow for me. But it works. Probably related to no negative for CFG1.
I haven't seen horizontal line artifacts (with the RL2 version). The most recent images I've posted, and the workflow image in the model description, have the ComfyUI workflow I recommend starting with. I'd double check that and see if any settings are off. If you still run into issues feel free to share gens with me and I can dig deeper.
@nutbutter Yea I have custom everything and using your model quanted in my custom quant, so it might either amplify issues, or even possibly create them. Once my tokens reset, I will dig a bit more into it.
But kudos for creating most prompt following version of FLUX 9B I saw so far. Like, in my case its nearly 100% of prompt followed.
I will try that WF and see what it does. Dont think its quant as regular FLUX quanted same way exhibits no issue.
@nutbutter Btw. batch size 8 in your ComfyUI workflow, are you nuts? Almost crashed my PC..
@nutbutter Found it, my issue. One of solvers I use is not suitable for your model for some reason. Will try to figure out why as its interesting.
For now, I wouldnt use DDIM from RES4LYF. :) And probably neither regular DDIM, no clue if that works anyway.
@Mescalamba haha, oops, good catch. Should update that for the average machine. I was testing on my training rig which has no problem with batch size 8
Take it from me, I've been in the ai scene since sd 1.4. and I've gone through a LOT of checkpoints. I've tried many lora and tested all models up to now. This one right here, If all goes well with training, this is the future of ai. Just try it yourself and see. Krea2 is just the new shiny toy, this is the diamond in the rough you've been searching for.
I think all us "old schoolers" have been waiting for Bigasp 3 to drop
Cant, PC is too old xD
Base is actually quite good too TBH with the right negatives. It likes higher CFG than normal Flux Klein 9B Base I find. Around 6.5 CFG / DPM++ 2S Ancestral / Simple scheduler / 28 steps seem to be good settings for it. Both Asp V3 Base and RL seem to be able to do normal hi-res-fix style upscaling just fine, same way as normal Klein 9B also. So no huge need to push the res up super high for the initial gen anyways.
base with this setup is phenomenal, worth the extra steps for sure
I think 20 steps in the default workflow might be slightly overshooting it for RL2 also BTW. It seems to often give good results that aren't usually as harsh-looking at as low as 8 steps in some cases.
ooowee finally something.. still use your SDXL model, still best model for trans so I was looking for a newer iteration from you and finally! woop woop
Has anyone had luck with Turbo lora's with it? I only have one turbo lora but it just messes everything up, not sure if turbo lora's just wont work with it period or there's a specific lora that might be needed with this?
Thanks Sir Nutbutter, I had actually been checking your page lately so see if you had posted anything new or if you;'d vanished lol
The latest version of the model, RL2, is a kind of "turbo", at least compared to Base. Base needs 40 step with CFG, so 80 model calls. RL2 works fine at 20 steps with no CFG, so 4x faster. But after RL is done I'll try training a real turbo version (something in the 4 step range).
Oh and I did a quick test on trans content and seems to work well, so let me know how it works for you and any feedback you have.
Because of this model I ended up removing around 50gb+ of models and loras. You have a real winner here and it's only "alpha". Keep up the fine work!