[Feature] Add reference-only preprocessor modes (reference_adain, reference_adain+attn)

[SolicTous]

Feature Summary

Add support for reference-only preprocessor modes to control image generation using a reference image

Detailed Description

The sd-webui-controlnet project has implemented reference-only preprocessing modes (reference_adain, reference_adain+attn) that allow controlling the output generation using a reference image.

Key features:

• Transfers style and visual characteristics from a reference image to the generated output
• Works without requiring additional ControlNet model files (uses the base SD model's attention mechanisms)
• Can be combined with traditional ControlNet models (canny, depth, pose) for more complex workflows

How it works:
These modes modify the self-attention layers during generation by injecting features from the reference image:

• reference_only: Direct substitution of K/V matrices from reference image
• reference_adain: Applies Adaptive Instance Normalization before attention injection
• reference_adain+attn: Hybrid approach combining both methods (recommended for style transfer)

References:

• Main: https://github.com/huggingface/diffusers/blob/main/examples/community/stable_diffusion_controlnet_reference.py
• Implementation discussion: [New Preprocessor] The "reference_adain" and "reference_adain+attn" are added Mikubill/sd-webui-controlnet#1280
• Usage examples: https://www.reddit.com/r/StableDiffusion/comments/13lla5c/a_deep_dive_into_the_new_reference_controlnets/

This feature would enable powerful style transfer and reference-based generation in stable-diffusion.cpp without additional model loading.

Alternatives you considered

No response

Additional context

No response