arXiv · 2026

⚡️FlashRender⚡️

Few-Step Generative Rendering via Camera-Controlled Video MeanFlow

1EverEx    2Yonsei University    3Korea University

†Corresponding authors

arXiv Paper Hugging Face Code
Teaser figure
TL;DR — Given an input video and a target camera trajectory, FlashRender generates a high-quality target retake within seconds.
25×

Lower sampling cost

NFE

4-NFE vs. 2×50-NFE baselines

+18%

Better camera controllability

TransErr / RotErr

than even multi-step baselines

+14%

Better geometric consistency

Dyn-MEt3R / MEt3R

than few-step baselines

+6.3%

Better visual quality on OOD

VBench

than even multi-step baselines

Motivation

Naive few-step sampling introduces discretization error, which manifests in generative rendering as sampling-step-dependent camera motion.

Camera motion changes with the sampling-step count

For the same scene and target camera trajectory, changing only the number of sampling steps can lead to substantially different realized camera trajectories.

ReCamMaster [ICCV'25]

2 steps
4 steps
8 steps
50 steps

ReDirector [CVPR'26]

2 steps
4 steps
8 steps
50 steps

GeoAlign [arXiv'26]

2 steps
4 steps
8 steps
50 steps

And distillation inherits the problem

Distill those models down to 4 steps, and performance degrades because their multi-step dynamics are difficult to approximate with only a few sampling steps.

ReDirector · 2×50 NFE
ReDirector-FS · 4 NFE
GeoAlign · 2×50 NFE
GeoAlign-FS · 4 NFE

FlashRender starts by enforcing the consistent camera control. We find that this significantly lowers denoising trajectory curvature, facilitating step distillation.

Method

FlashRender is fine-tuned from Wan2.1-1.3B-CamCtrl in three stages, each designed to address a distinct challenge in few-step generative rendering.

Architecture overview
1

RETA

Stabilize camera control across timesteps.Reduce denoising trajectory curvature.

Aligns intermediate source-video representations with target-view features from a frozen VGGT encoder. This internalizes the geometric transformation, enabling consistent camera control and reducing the trajectory curvature.

2

MeanFlow

Shortcut the full denoising trajectory.Mitigate discretization error.

Building on the lower-curvature denoising trajectory induced by RETA, we fine-tune the multi-step model with the MeanFlow objective to learn average velocity fields that shortcut the full denoising trajectory, effectively mitigating discretization error.

3

On-policy flow map distillation

Address self-rollout errors.Optimize under fixed few-step sampling.

The MeanFlow model is rolled out using the same four-step schedule as inference and optimized on its own samples with DMD and an adversarial objective, correcting self-rollout errors caused by the training–inference mismatch.

Analysis on RETA

With RETA, the realized camera trajectory remains consistent across sampling steps.

FlashRender-MS (Stage 1) · with RETA

2 NFE
4 NFE
8 NFE
50 NFE

RETA flattens the denoising trajectory

A straighter denoising trajectory is easier to approximate with only a few steps, facilitating subsequent step distillation. RETA consistently lowers the curvature throughout the schedule, achieving the lowest curvature along all baselines.

C(ti) = ‖ (ẑi − ẑi−1) / (ti − ti−1) − (X1 − X0) ‖22 — how far each denoising step veers from the straight line between noise X1 and data X0, squared. A straighter denoising trajectory is easier to approximate with only a few steps.

Lower-curvature trajectory makes MeanFlow training easier

RETA enables the MeanFlow model to more accurately shortcut the full denoising trajectory and effectively mitigate discretization error during few-step sampling.

Accurate trajectory shortcutting enables better on-policy distillation

With RETA, video quality improves consistently across all three stages. More accurate trajectory shortcutting in turn enables better on-policy distillation.

without RETA

Stage 1 · 4 steps
Stage 2 · 4 steps
Stage 3 · 4 steps

with RETA

Stage 1 · 4 steps
Stage 2 · 4 steps
Stage 3 · 4 steps

Comparisons

Results on the DAVIS dataset — 50 source videos × 10 target trajectories.

Method Visual Quality Geometric Consistency Camera Accuracy
Aesthetic ↑ Imaging ↑ Dyn-MEt3R ↑ MEt3R ↓ TransErr ↓ RotErr ↓
Multi-step (2×50-NFE)
CogNVS0.21600.43000.68450.40360.076810.878
TrajectoryCrafter0.50460.60710.73380.32720.06979.115
Vista4D0.50950.65270.78120.31230.02232.371
GCD0.39980.49280.68980.44380.106222.853
ReCamMaster0.50640.64610.78570.34720.02922.347
ReDirector0.51490.66680.84770.30730.01651.666
GeoAlign0.52500.66830.85320.31110.01491.495
FlashRender-MS0.52140.66250.84910.30590.01431.486
Few-step (4-NFE)
NeoVerse0.50120.67150.75230.36190.03433.495
ReDirector-FS†0.50380.64370.83510.34920.01791.746
GeoAlign-FS†0.50830.64780.84650.31610.01581.528
FlashRender0.51820.66540.85710.30710.01221.236

Best per column within each group (multi-step / few-step) in bold. FS† denotes few-step baselines built by applying our MeanFlow-based on-policy flow map distillation.

Wall-clock time to render one video, measured for the video generative model only.

Baseline comparisons

FlashRender and NeoVerse run at 4-NFE; other baselines use 2×50-NFE.

Input Video
TrajectoryCrafter
CogNVS
NeoVerse
ReCamMaster
ReDirector
GeoAlign
FlashRender (Ours)

Few-step comparisons

All models use 4 sampling steps, while ReDirector and GeoAlign use double the NFE due to CFG guidance. FS† uses our MeanFlow-based on-policy distillation.

Sampled directly at 4 steps

Input Video
ReDirector
GeoAlign

Distilled to 4 steps with our method

ReDirector-FS†
GeoAlign-FS†
FlashRender (Ours)

More results

The same input video retaken by FlashRender following ten target trajectories of the evaluation protocol. Trajectory and scene are chosen independently.

Input Video
FlashRender (Ours) · Pan Right

BibTeX

@article{park2026flashrender,
  title   = {FlashRender: Few-Step Generative Rendering via
             Camera-Controlled Video MeanFlow},
  author  = {Park, Byeongjun and Kim, Byung-Hoon and Chung, Hyungjin},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}