mewvan/renderflow-mini-diffusion

RenderFlow-mini — DDPM noise-diffusion baseline (run B)

Baseline arm for the RenderFlow scaled reproduction: identical UNet, data, training budget and G-buffer conditioning as run A (mewvan/renderflow-mini-bridge), but trained as an ε-prediction DDPM from Gaussian noise (cosine schedule) and sampled iteratively with 50 ancestral steps — the "iterative diffusion" paradigm RenderFlow argues against.

Test results (300 frames):

model	inference	PSNR	SSIM	LPIPS	latency/frame
A: bridge (albedo prior)	1 step	25.50 dB	0.796	0.230	20.1 ms
B: diffusion (noise prior)	50 steps	13.47 dB	0.501	0.536	1007.6 ms

B's PSNR varies ±0.50 dB across sampling seeds (stochastic); A varies ±0.16 dB (deterministic). Note: at equal training budget the diffusion baseline is undertrained relative to the paper's baselines (RGB-X was fine-tuned 100k steps); the gap here therefore overstates the paper's ~10×/SOTA claim, but confirms the direction: for G-buffer→image maps, an albedo-prior single-step flow beats a noise-prior iterative diffusion trained on the same compute.

Strengthen the baseline