ReChannel: Extracting Geometry from Generative Models

The ReChannel project uses LoRA adapters on top of the FLUX.2 architecture to extract depth maps, normals, and segmentation from a single RGB frame.

Compiled by Sergey KostenchukPublished 2026-07-14Updated 2026-07-14

2026-07-14 Research HuggingFace

📐 ReChannel: Extracting Geometry from Generative Models

The ReChannel project enables the extraction of depth maps, surface normals, masks, and segmentation from a single RGB frame. The method utilizes lightweight LoRA adapters (~33K parameters) on top of a frozen FLUX.2-klein-base-4B architecture, transforming generation into a process of dense spatial data readout.

🌍 This revolutionary approach turns multimodal generative models into high-precision computer vision tools. Using a frozen DiT (Diffusion Transformer) with a minimal number of trainable parameters makes the task of geometry extraction extremely efficient.

👤 It is now possible to use powerful models like FLUX for computer vision tasks without training massive neural networks from scratch, by simply adding tiny adapters.

Source 1: https://github.com/xmz111/ReChannel Source 2: https://arxiv.org/abs/2607.06553

Sources