FlashRender Retakes Source Video at 25x Lower Sampling Cost

FlashRender, a video re-rendering framework from researchers at EverEx, Yonsei University and Korea University, claims to match multi-step baselines in visual quality and geometric consistency using 25x fewer sampling steps, cutting the cost of retaking a source video along a new camera path to 4-NFE against a 2×50-NFE reference.

The technique bundles three pieces the authors argue are complementary. Representation Transformation and Alignment (RETA) aligns hidden source-video features with target-video features pulled from a frozen visual geometry model, which the paper says "directly encodes the geometric transformation within the source-video stream, enabling sampling-step-consistent camera control." The model is then fine-tuned with the MeanFlow objective on the lower-curvature denoising trajectory RETA induces, and distilled with on-policy flow maps to correct self-rollout errors under fixed few-step sampling.

On DAVIS at 4-NFE, FlashRender reports the best few-step numbers against ReDirector, GeoAlign, NeoVerse, ReCamMaster, TrajectoryCrafter and CogNVS, with +18% on camera controllability, +14% on geometric consistency and +6.3% on visual quality on out-of-distribution camera paths, the regime where prior camera-controlled generators typically fall over. The abstract puts the topline as matching multi-step baselines "at 25x lower sampling cost while achieving superior camera controllability, even under out-of-distribution target camera trajectories."

The project page publishes the benchmark tables but does not specify wall-clock latency, GPU class, or a code release.