FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow
FlashRender retakes a source video along a new camera path in seconds while cutting sampling cost by 25x.
The paper says existing multi-step generative rendering models suffer from sampling-step-dependent camera control, a discretization error that bends the denoising trajectory. FlashRender adds RETA to align source-video hidden representations with target-video geometry features, then applies MeanFlow fine-tuning and on-policy flow map distillation. In experiments, the combined system matches multi-step baselines on video quality and geometric consistency while improving camera controllability, including on out-of-distribution camera trajectories. HF Daily Papers' note
The paper says existing multi-step generative rendering models suffer from sampling-step-dependent camera control, a discretization error that bends the denoising trajectory. FlashRender adds RETA to align source-video hidden representations with target-video geometry features, then applies MeanFlow fine-tuning and on-policy flow map distillation. In experiments, the combined system matches multi-step baselines on video quality and geometric consistency while improving camera controllability, including on out-of-distribution camera trajectories. HF Daily Papers' note
score 4