Researchers have unveiled FlashRender, a novel framework for few-step generative rendering that can retake a source video along a target camera trajectory in seconds. The method addresses a key limitation in existing multi-step generative rendering models: sampling-step-dependent camera control, which stems from discretization error.
The team identified that resolving this inconsistency significantly reduces the curvature of the denoising trajectory, making it easier to distill the model into fewer steps. To achieve this, they introduced Representation Transformation and Alignment (RETA), which aligns hidden source-video representations with target-video features from a frozen visual geometry model, directly encoding geometric transformation in the source-video stream.
Building on RETA, the model is fine-tuned with a MeanFlow objective on the lower-curvature denoising trajectory, enabling it to better handle discretization error. Finally, on-policy flow map distillation corrects self-rollout errors under fixed few-step sampling.
Extensive experiments show that RETA, MeanFlow, and on-policy flow map distillation play complementary roles. Together, they allow FlashRender to match multi-step baselines in video quality and geometric consistency at 25x lower sampling cost, while achieving superior camera controllability, even for out-of-distribution target camera trajectories.