mewvan/renderflow-mini-keyframe

RenderFlow-mini — keyframe guidance adapter (run E)

Stage-2 adaptation for the RenderFlow scaled reproduction: the run-A forward model (mewvan/renderflow-mini-bridge) is frozen; a cross-attention adapter injects the ground-truth keyframe (frame 0 of each 5-frame sequence) into the UNet at ≤64×64 resolution (higher resolutions OOM'd on 24 GB — see report caveats).

Test results (300 frames):

variant	PSNR	SSIM	LPIPS
without keyframe (adapter bypassed)	25.03 dB	0.762	0.240
with keyframe	24.97 dB	0.791	0.309

Finding: not validated at this scale. SSIM improves (+0.029) but PSNR is flat and LPIPS worsens; all deltas are within the ~0.2–0.5 dB kernel-noise floor observed between identical eval runs. Plausible causes: 5-frame clips make the keyframe near-redundant, the adapter is far smaller than the paper's (plus we dropped the paper's RoPE temporal encoding and the LoRA-in-FFN variant), and 6k adapter steps may under-train it. The paper's keyframe-gap ablation (13–49 frames) operates in a regime our 5-frame clips cannot reach.

Improve the adapter