What parameters are relevant for Flux2-Klein image2image?

[ddpasa]

I am doing some basic image2image with flux2-klein model, an example command is below:

sd-cli --diffusion-model /models/flux2_klein_9b/flux-2-klein-9b-Q8_0.gguf --vae /models/flux2_klein_9b/ae.safetensors --llm /models/flux2_klein_9b/Qwen3-8B-Q8_0.gguf --steps 4 -v --diffusion-fa --vae-tiling --vae-conv-direct -r "/input/testimg.jpg" --cfg-scale 1.0 -p "preserve exact body proportions, only enhance quality, ultra-detailed, photorealistic, sharp focus, clean textures, professional photography, natural lighting, high resolution, film grain subtle" --negative-prompt "blurry, pixelated, jpeg artifacts, compression noise, oversmoothed, plastic, low quality, deformed, watermark" --output "/output/testimg_flux2k_20260518_1018.jpg" --seed 978904

I want to control the image2image strength. Normally with older SDXL models you would do this with the --strength parameter. However, it does not seem to do anything here, as I get the same output regardless of the value (tried between 0 and 1).

Which parameters are relevant for flux2, especially for image2image?

[stduhpf]

--strength only works for images passed with -i (--init-img), not -r (--ref-image)

There is the --img-cfg-scale, wich I don't quite remember if it is actually applied to edit models, or it's just for inpaint models, you could give it a try.

[ddpasa]

--strength only works for images passed with -i (--init-img), not -r (--ref-image)

There is the --img-cfg-scale, wich I don't quite remember if it is actually applied to edit models, or it's just for inpaint models, you could give it a try.

Thanks @stduhpf ! I guess I don't understand the difference between -r and -i. I get nice img2img behaviour with -r, but it ignores --strength. With -i, the results are responsive to --strength, but usually very bad.

[stduhpf]

-i controls the initial image. -r adds a reference image

When the initial image is set, instad of starting denoising from pure noise, the model starts off from the initial image with some noise added to it (controlled with the strength parameter). Every diffusion and flow model supports that, because that's how they are trained.

The reference images, on the other hand, stay clean without noise, the model can access them directly to transfer their content to the output. Only a few models support that so far.

[ddpasa]

Ok, then I have been doing image2image wrong, but getting good results nevertheless :) Surprising that I get better results with -r than -i