MVID — multi-view intrinsic decomposition

Weights for the MVID demo. Decomposes images into I = A · S + R:

• A — diffuse albedo (reflectance, not base colour; metals go dark)
• S — three-channel coloured irradiance
• R — non-diffuse residual: speculars, light sources, interreflections

VGGT-Omega backbone (1.35B params) with alternating frame-wise and global attention, so a set of views is decomposed jointly rather than per-image.

Files

file	dtype	size
mvid_ep186_bf16.safetensors	bfloat16	2.5 GB

Load into MVIDModel with strict=False. Inference runs under torch.autocast(bfloat16); against the fp32 weights the outputs differ by <0.002 mean on albedo/shading/residual/depth.

Trained at 512 px. Inputs must be a multiple of patch size 16 — the aggregator silently drops the remainder otherwise.