Flux De-distilled Training Samples Garbled

[kmacmcfarlane]

Should I expect sample prompts for de-distilled flux models to be working yet?

On the sd3 branch, when finetuning a de-distilled model (nyanko7/flux-dev-de-distill) the sample image output does not seem to be working correctly. Inference output from the training checkpoints in ComfyUI look correct when applying CFG.

Example prompt.toml

[prompt]
#negative_prompt = "makeup, nude, naked, border, frame, signature, text, watermark"
width = 1024
height = 1024
scale = 1
sample_steps = 60
seed = 420

[[prompt.subset]]
prompt = "my cool prompt here"

I also tried with scale set to 3.0 thinking the inference code might need the guidance scale set higher for de-distilled inference. That resulted in the same issue.

[kohya-ss]

Currently, CFG cannot be applied to sample image generation. We will support it soon.