style transfer flux3

[KintCark]

How do I do style transfer with flux 3 klein?

I trying but I'm not getting the results I want here is my command line: ./bin/sd-cli --diffusion-model ~/stable-diffusion.cpp/build/models/flux-3-image-klein-Q8_0.gguf --llm ~/stable-diffusion.cpp/build/models/Qwen3VL-4B-Instruct-Q8_0.gguf --vae ~/stable-diffusion.cpp/build/models/ae.safetensors --llm_vision ~/stable-diffusion.cpp/build/models/mmproj-Qwen3VL-4B-Instruct-F16.gguf --cfg-scale 1.0 -p "Transform the artistic style of this image into A close-up, low-angle shot of a figure standing amidst a chaotic, vibrant environment. The scene is a fusion of raw, distressed punk textures—torn fabrics, exposed wiring, and neon signs—but the overall atmosphere is one of contained, luminous safety. The lighting is high-contrast, utilizing deep shadows juxtaposed with intense, saturated neon glows. The composition uses a strong diagonal line to lead the eye through the cluttered foreground, while the background suggests a controlled, dimly lit, safe underground space. Materiality emphasizes wet asphalt, frayed denim, and flickering electrical light. Mood is electric, defiant, yet deeply grounded and resilient. Wide-angle lens to capture the dynamic tension." -v -H 160 -W 240 --steps 4 --threads 6 --strength 0.95 -r ~/stable-diffusion.cpp/build/models/Astro-Bio.png -o ~/stable-diffusion.cpp/build/models/Transfer.png

What am I doing wrong

[ngo-binh0108]

A few things in your command work against you:

1. Steps: the distilled Klein model is meant for 16 steps (docs/flux3.md), not 4. At 4 steps it can't move far from the reference.

2. Size: -H 160 -W 240 is very small. There's almost no room for texture, lighting or brushwork, which is what a style is. Try at least 768-1024 on the long side, keeping the reference's aspect ratio.

3. --strength does nothing here. It only applies to img2img with -i (see src/pipeline/image.cpp). With -r, the image is passed as a reference to the model, and how much it changes is controlled by the prompt.

4. Prompt: with -r, Klein works like an editing model and expects an instruction, not a scene description. Your prompt describes a new scene, so the model has to choose between your image and the text. Keep it short and say what to keep, for example:

   Repaint this image in a gritty neon punk style: high-contrast lighting, saturated neon glows, distressed textures. Keep the composition and the subject.
   

Something like:

./bin/sd-cli --diffusion-model .../flux-3-image-klein-Q8_0.gguf --llm .../Qwen3VL-4B-Instruct-Q8_0.gguf --vae .../ae.safetensors -r .../Astro-Bio.png -p "Repaint this image in a gritty neon punk style: ... Keep the composition and the subject." --cfg-scale 1.0 --steps 16 --sampling-method euler -W 1024 -H 704 -v -o .../Transfer.png

(--llm_vision isn't in the FLUX.3 examples in docs/flux3.md, so you can leave it out.)

If the result still copies the reference too closely, try the Klein Base model with --cfg-scale 3.5 --steps 20. CFG gives the prompt more weight.