V1 - 6/23 - Initial Release - GGUF'S ARE HERE!
===========================================================================
CRUX (chroma/flux) hybrid models are created by DR34MSC4PE with the best of both worlds:
In a blurb? It is a hybrid model between Chroma/Flux that puts back the removed components (modulation layers, etc) from FLUX.1-dev by selectively merging it with a special variant of C4PACITOR but, importantly, keeping Chroma's de-distilled (unchained) nature and output layers.
Can be used with Chroma or FLUX de-distilled workflows (CFG enabled)
Can even be used with flux.1-dev workflows, but you'll need to increase the FluxGuidance to > 4 for adherence most times. Sampler research into "what all works" is ongoing but has been surprising!
For the moment, it is recommended to use CHROMA or classic CFG enabled workflows (including negative prompts!) vs. FLUX unless you want to help research the boundaries with us!
Specific anime/illustration tuning to introduce some lost artistic concepts
NSFW tuned and capable of realism and artistic/anime images. Female anatomy is well represented as a result of CHROMA + C4PACITOR's advance knowledge on these subjects. It tends to lean toward the latter's style.
FLUX LoRA seem to work but testing is ongoing for efficacy.
We have not tested this in any inference tool except ComfyUI yet - YMMV! Keep us posted!
Like our work? Buy us a coffee: https://ko-fi.com/dr34msc4pe
===========================================================================
DR34MSC4PE is
c0ur4ge (training/qa/inference code) /
eraser851 (training/captions/tooling code/data/qa)
===========================================================================
Description
The GGUF versions of CRUX-v1
FAQ
Comments (6)
Forge Sampler-Scheduler Combinations
I've been doing some tests on Forge combinations for CRUX. Here are the results. Generally, I like the DPM and Heun based combinations best, with distilled cfg 0, and cfg around 3.5, but try values up and down from there, as high or low as you like:
DPM++ 2M Beta or Simple
DPM2 Beta or Simple
DPM2a Beta or Simple
Heun Beta or Simple
HeunPP2 Beta or Simple
Euler Beta or Simple
Euler A Beta or Simple
IPNDM Beta or Simple
DEIS Simple
LMS Simple
IPNDM_V Simple
Most other combinations in Forge will either get you nothing, blurry images, or a burned image. Some combinations may work with very high steps. My tests were done with 32 steps, to keep it in the middle ground.
Please note that you may need to carefully use negative prompts to clear up anatomy concerns. You may also want to increase steps for a final generation, and you will likely want to try smaller values (about 3/4 of recommended values) on LoRAs.
This test was done with a fairly simple prompt and the same seed. A simple negative prompt was also included.
Oh, and sometimes a useful tool takes more than a minute to master it, but when you do, you'll find yourself surprised with what you can do with it. That's why I use DR34MSC4PE models. ;)
I hope someone finds this useful.
where can i find a list of available art styles for this model? thanks
Basically - prompt for them! We largely have datasets that also introduce famous art styles and small amounts of “well know art” as well but frankly, since this is a pretty wild merge of two fairly divergent but capable models, there’s still a bit of discovery to be had here.
Really nice
If you've tested on comfyui, then please share workflow? I do have Flux workflows, but I'm curious about yours.
Try the stock Chroma (or any chroma) txt2img workflow! :)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.