FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow

FlashRender. Few-Step Generative Rendering via Camera-Controlled Video MeanFlow. By Byeongjun Park and two others, from EverEx, Yonsei University, and one other institution. Published on the third of September. Its code repository is worked out: about three hundred source files, with a readme and a licence.

Generative rendering takes a source video and retakes it along a different camera trajectory. Doing that well currently takes many sampling steps, and cutting the step count is where the quality goes. The paper starts from a specific diagnosis rather than a general complaint about speed: in existing multi-step models, camera control depends on how many sampling steps you use. That sampling-step-dependence is a prominent manifestation of discretization error, and it is the thing standing between these models and a fast version of themselves.

The insight is that this inconsistency is not merely a symptom. Resolving it substantially lowers the curvature of the denoising trajectory, and a lower-curvature trajectory is much easier to distil into few steps. So fixing camera consistency first is what makes the subsequent speedup possible, rather than being a separate quality concern.

Three pieces, applied in order. Representation Transformation and Alignment lines up hidden source-video representations with target-video features taken from a frozen visual geometry model, which encodes the geometric transformation directly inside the source-video stream and gives camera control that no longer varies with step count. The model is then fine-tuned with a MeanFlow objective on the lower-curvature trajectory that alignment produced. Finally, on-policy flow map distillation corrects the self-rollout errors that appear once sampling is fixed at a few steps. The authors report that all three play complementary roles, which is the ablation the argument needs.

The result matches multi-step baselines on video quality and geometric consistency at twenty-five times lower sampling cost, and achieves better camera controllability than they do, including under out-of-distribution target trajectories. Better control rather than merely equal control is the notable part, since a distilled model would normally be expected to give some up.

The transferable shape is the ordering. A defect that looks like a quality problem turned out to be the obstacle to compression, and repairing it first was what made the model distillable at all.

FlashRender. Few-Step Generative Rendering via Camera-Controlled Video MeanFlow. By Byeongjun Park and two others, from EverEx, Yonsei University, and one other institution. Published on the third of September. This is the long version: the whole paper rather than the abstract.

Generative rendering takes an existing video and re-renders it along a camera trajectory you specify. It can synthesize dynamic novel views under occlusions and in regions never filmed, which supports capturing footage from physically impractical viewpoints and stabilizing shaky material. The obstacle to using it is mundane: existing methods need many sampling steps, so inference is expensive.

The obvious remedy is to use fewer steps, and coarse sampling introduces discretization errors that become pronounced along highly curved denoising trajectories. That much is general knowledge. What this paper contributes first is a specific diagnosis of how that error shows up here, and the diagnosis is what everything else follows from. They identify sampling-step-dependent camera control as the key failure mode. Given an identical target camera trajectory, changing the number of sampling steps alters the camera motion actually realized in the generated video, which changes scene scale and leaves dynamic objects inconsistently grounded in space.

Which sets up the paper's actual insight: resolving that inconsistency substantially lowers the curvature of the denoising trajectory, and a lower-curvature trajectory is much easier to distil into few steps. So fixing camera consistency is not a separate quality improvement applied alongside the speedup. It is what makes the speedup possible. Three components then apply in order. Representation Transformation and Alignment aligns intermediate source-video representations with target-video features taken from a frozen visual geometry model, which directly transforms the source representations toward the target view and supplies timestep-independent target-view geometry. That is what removes the step dependence: the geometric transformation is encoded in the source-video stream itself rather than being re-established at each denoising timestep.

The model is then fine-tuned with a MeanFlow objective on the lower-curvature trajectory that alignment produced, so the objective aimed at discretization error is applied to a trajectory where discretization error is already smaller. Finally, on-policy flow map distillation corrects the self-rollout errors that appear once sampling is fixed at a few steps, which is a different class of error again: not the trajectory being curved, but the model compounding its own mistakes when it runs without supervision. The authors report that all three play complementary roles, which is the ablation the argument requires, since the sequential story would collapse if any one of them accounted for most of the gain.

The reported result is matching multi-step baselines on both video quality and geometric consistency at twenty-five times lower sampling cost, with better camera controllability than they achieve, including on target trajectories outside the training distribution. Two things about that are worth separating. Matching on quality is the expected shape of a successful distillation, and it is the weaker half. Beating the baselines on control while running twenty-five times cheaper is the half that tests the diagnosis, because it is what the step-dependence account predicts and what a generic speedup would not produce.

The limitation is stated plainly and it is a real one. Performance degrades substantially as the sampling step count is reduced further. At two function evaluations the MeanFlow model becomes less effective at mitigating discretization error under extremely low sampling budgets, which in turn limits the effectiveness of the subsequent distillation stage. So the method has a floor, and the chain of three components fails from the middle outward when you push past it. The authors also note their model still underperforms recent four-dimensional reconstruction methods at localizing dynamic objects under moving input cameras, and suggest stronger scene-reconstruction backbones as a route to better geometric accuracy there.

The transferable shape is the ordering, and it generalises well beyond video. A defect that looks like a quality problem turned out to be the obstacle to compression, because it made the thing being compressed unstable. Repairing it first was what made the model distillable at all, and the speedup came as a consequence rather than as the goal pursued directly.