img2img not working

[InformEthics]

Thank you for doing this great project!
using mlx-FLUX.1-schnell-4bit-quantized model and providing a --image-path argument the img2img function does not work. Output image is the same as input image:

diffusionkit-cli --prompt "Hand drawn illustration of a boy reading for his teddy bear sitting on a treehouse, the boy is wearing a hat with bear ears, a scarf, and rainboots, next to him is a bookshelf with books on it, in the background it is raining, autumn colours." --image-path ~/Desktop/sketch.png --model-version argmaxinc/mlx-FLUX.1-schnell-4bit-quantized --denoise 0.5 --height 1360 --width 768 --step 4 --output ~/Desktop/test2.png

[mplawner]

Hi @InformEthics

I’ve noticed something similar while experimenting and thought it might help shed some light on what’s going on. It seems like the --denoise and --steps settings interact in a way that affects the actual number of steps performed during processing.

For instance, when I used --denoise 0.5 with --steps 4, the system only ran 2 actual steps instead of 4. From what I can tell, it looks like the code might be multiplying the --denoise value by --steps to determine the effective number of steps. So, 0.5 × 4 = 2 steps in that case.

When I bumped up the --denoise to 0.775 while keeping --steps at 4, the output actually went through all 4 steps. My guess is that the calculation (0.775 × 4 = 3.1) gets rounded up to 4 steps, which seems to give more refined and expected results.

If you’re seeing similar behavior, try adjusting the --denoise value—something in the range of 0.7–0.8 worked well for me. It allowed the system to process the full number of steps and produce more detailed outputs.

I modified the prompt a bit, and got the attached result.

diffusionkit-cli --prompt "Hand drawn illustration of a boy reading for his teddy bear sitting on a treehouse, the boy is wearing a hat with bear ears, a scarf, and rainboots. His teddy bear is cuddling up to him. A lantern with butterflys hangs from the awning. Next to him is a bookshelf with books on it, in the background it is raining, autumn colours." --image-path sketch.png --model-version argmaxinc/mlx-FLUX.1-schnell-4bit-quantized --height 1360 --width 768 --step 4 --output test2.png --verbose --denoise 0.775
(Seed was 1736614223)

Hopefully, this helps!