Based on my Centerfold Flux 5, trying to merge in Chroma and Krea.
Diffuser version 2.3 https://huggingface.co/Tiwaz/CenKreChro_V2_3
For version 2.3 I will add the quantizations under the same "tab".
Big thanks to Spooknik for making the SVDQ versions https://huggingface.co/spooknik/CenKreChro-SVDQ
Q4 KM works fine in Comfy UI but not in Forge for example.
If you need the diffuser version: https://huggingface.co/Tiwaz/CenKreChro (not 100% tested yet).
You can try using it as low as 12 steps even if I prefer 30.
Description
FAQ
Comments (16)
Issue with levels of black. I'm post it here not to clutter my other issue. Difference between pure flux-krea-dev vs CenKreChro, where CenKreChro produces gray artifacts compare to pure flux-dev model and flux-dev+krea-lora-rank-256 on lower weight.
CenKreChro - gray grid artifacts, not true black. Clearly visible when increasing gamma level.
I compared levels of black of CenKreChro vs flux-dev vs flux-krea-dev vs flux-dev + krea-lora-rank-256 here vs CenKreChro FP8:
https://imgur.com/a/TCryxdP
Nice work! Looks promising. Any chance to release the FP16? Also does this works well with LoRAs?
To be honest did not take notes when I made it and never made it as FP16, so probably not.
Will I make a successor as FP16 from scratch yes maybe, I have some ideas.
I did not Lora test it personally so can't comment on that. But I imagine FLux / Krea Loras should be fine.
@TiwazM Fair enough. Thank you for quick responding. Hope you get a greater iteration.
Great stuff, and the Nunchaku version is the cherry on top. Big thanks!
PS: What's your favorite sampler/scheduler combo for it?
I can't decide at the moment between, euler with simple or beta or kl_optimal, sometimes I like one or the other better. Res2s creates the banding like I have seen in some other models.
Fibonnaci or kl_optimal and res_multistep ancestral with https://github.com/obisin/ComfyUI-FSampler or beta_1_1 from comfyppm.
Flux do love it's ugly people tho.
It's working at 10 steps for me in SVDQ-int4. Didn't try anything risque, yet but the speed is very promising. The chroma SVDQ we never got.
What is exactly for use the checkpoint loader? because other loader checkpoint unet loader doesnt works: 'nonetype' object has np attribute 'model'
For the SCDQ FP4 or INT4? Either Nunchaku in ComfyUI or Forge Classic Neo
Hello, your model is excellent and produces for me way more realistic images even with the nunchaku quant without any loras than chroma variants alone and is not even at all finicky with producing consistent results compared to pure chroma. however i have noticed one strange thing and am not sure if this is the part of training data that I cannot fix or something else, the model doesnt seem to understand the concept of a swimsuit with thick/wide shoulder straps. it always generates the straps as thin laces or just thin straps and no matter what controlnet I use, combinations of various controlnets, depth/canny/hed.... it never produces the same swimsuit with wide straps that are present on reference images. everything else down to face details is reproduced precisely... any idea how I would go about fixing this issue?
I really like this model! It produces better output than original Krea. Are you able to upload the FP16 precision version? That would be really cool! I will upload some images to the gallery soon. Thanks!
The author has made them available here - https://huggingface.co/Tiwaz/CenKreChro/blob/main/CenKreChro.safetensors
@Spooknik377 Thanks. I wasn't able to find it, though. The link you've provided leads to the FP8 model. FP16 Flux models are typically ~22 GB in size.
@mmdd2543 My bad, the FP16 is in the transformers folder. https://huggingface.co/Tiwaz/CenKreChro/tree/main/transformer
@Spooknik377 yes but not a real fp16, it is just the fp8 converted to diffusers