Self-consistency boosts video generation accuracy for visual reasoning tasks
Learning via Self-Consistency for Diffusion-based Video Reasoning
Computer Vision and Pattern Recognition
Summary
Video-based models that create or analyze videos often struggle because the process is random while the tasks need precise, consistent answers. The authors show that by generating multiple videos and combining their answers, they can get better, more reliable results. They also introduce a training method to teach the model to give better answers with just one attempt, saving time. Their techniques greatly improve performance in tasks like solving mazes or finding objects in videos without needing extra labeled data.
What this means in practice
- •For computer vision engineers: Improve accuracy in video-based perception tasks like object tracking or segmentation by aggregating multiple model outputs for a consensus answer.
- •For machine learning practitioners: Reduce inference time while maintaining accuracy by using early readouts and training the model to internalize consensus predictions for single-rollout results.
Authors
Zhenghao Ni, Weimin Qiu, Meng Tang
Abstract
Video generation models have demonstrated emerging zero-shot capabilities for visual reasoning, perception, and other vision tasks. However, diffusion-based video generation is inherently stochastic, while many downstream vision tasks are deterministic. Motivated by the effectiveness of self-consistency in chain-of-thought reasoning for large language models, we investigate whether self-consistency can similarly improve diffusion-based video reasoning. We first introduce a training-free test-time scaling method that samples multiple video generations and aggregates their predictions through self-consistency. Specifically, we aggregate extracted paths, locations, or masks from multiple rollouts into a consensus prediction. To reduce the inference overhead of multi-rollout generation, we read out predictions early in the denoising trajectory, which preserves consensus quality while reducing denoising steps by more than half. We further propose Rejection Fine-Tuning (RFT) to distill consensus predictions into the video generation model. The resulting model internalizes the benefit of multi-sample consensus and requires only a single generation at inference time, while substantially outperforming the original model. Experiments on three tasks, including maze solving, visual search, and referring segmentation, show that both our self-consistency inference and consensus distillation dramatically improve video-based perception and reasoning, without requiring ground-truth videos or task-specific verification. For visual search, self-consistency raises task accuracy from 48.4% for a single generation to 99.0%. The distilled model retains much of the consensus benefit with a single rollout. For 4-by-4 maze solving, consensus-based training improves the single-generation strict success rate from 72.0% to 84.0% with the same inference latency.