Experiment with chroma, turbo version with my small additions.
Run: Euler/Beta, CFG 1, Step 6-8
Text Encoder: https://huggingface.co/fmoraes2k/t5xxl_fp8_e4m3fn.safetensors/tree/main
Description
fp8_e4m3fn
FAQ
Comments (11)
Great to see a new Chroma Model! Is it a trained model, fine tuned, or a merge of Loras? What does it do better than the original Chroma?
This is a merge of the model with some loras in certain proportions, an attempt to create a normal turbo version.
Its not working for me, could you send me the workflow for ComfyUI, Bro.
@toya_san Thanks Bro
@toya_san could you please share the workflow again?
@bien622 It's the standard workflow found in Comfyui. Just follow the settings from the model's description.
Unbelievably.... it is even better then Z-Image-Turbo. With few LoRA and tiled sampling... i just can't believe that.
Definitely - you should proceed farther improvement of "experiment". Model need to recalibrate/rescale. 2-d pass almost every time output "spikes", what makes "inpaint/upscale" technique almost impossible.
Looking forward to your experimentation. Great job, man.
Dont work in DrawThings App.
Can you share fp16 version?
This is amazingly fast, but it has a hard time staying photo real.