mewvan/renderflow-mini-bridge

RenderFlow-mini — bridge matching, single step (run A)

Scaled-down reproduction of RenderFlow: Single-Step Neural Rendering via Flow Matching (Zhang et al., CVPR 2026).

Conditional UNet (~22M params, 256×256, pixel space), trained with Brownian-bridge matching (σ=0.005, discrete 4-step schedule ∪ uniform-t mixture), flowing albedo → fully shaded image, conditioned on albedo + normals + depth + tonemapped HDRI (FiLM).

Test results (300 frames, held-out procedural Mitsuba scenes, seed 0):

inference	PSNR	SSIM	LPIPS	latency/frame
1 step	25.50 dB	0.796	0.230	20.1 ms (a10g)
4 steps (Euler)	25.66 dB	0.810	0.231	78.4 ms

Determinism: ±0.16 dB PSNR across repeated identical runs (GPU kernel noise; no sampling stochasticity).

Trained: 15,000 micro-steps, effective batch 32, AdamW lr 1e-4, bf16 autocast, + LPIPS/gradient pixel losses (λ=1). Data: mewvan/renderflow-mini-dataset · Report: mewvan/renderflow-mini bucket README.

Extend this model