RenderFlow: Single-Step Neural Rendering via Flow Matching
Paper • 2601.06928 • Published
RenderFlow-mini — ODE flow matching ablation (run C)
Ablation arm for the RenderFlow scaled reproduction: identical UNet, data, budget and pixel losses as run A (mewvan/renderflow-mini-bridge), but trained with plain rectified flow / ODE (uniform t, no bridge noise).
Test results (300 frames, single-step inference):
| inference | PSNR | SSIM | LPIPS | latency/frame |
|---|---|---|---|---|
| 1 step | 25.19 dB | 0.778 | 0.231 | 20.1 ms |
| 4 steps (Euler) | 25.26 dB | 0.786 | 0.237 | 78.5 ms |
Run A (bridge/SDE) beats it by +0.31 dB PSNR / +0.019 SSIM at identical cost — same direction as the paper's Table 3.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support