Towards Subject Consistency over Dynamic Subject Sets in Video Generation

Subject-driven video generation (SDV-Gen) aims to produce videos of a specific subject by adapting a pretrained video model, enabling personalized and application-driven content creation. To achieve this goal, per-subject tuning methods require approximately 200 A100 GPU hours to generate a customized video, whereas zero-shot methods avoid per-subject tuning but typically rely on millions of subject-video pairs for the supervision, incurring massive network fine-tuning costs (10K-200K A100 GPU hours). We propose a data- and compute-efficient zero-shot SDV-Gen framework that avoids test-time per-subject tuning and the use of large-scale subject-video pairs. Our key idea decomposes SDV-Gen into (i) identity injection learned from subject-image pairs and (ii) motion-awareness preservation maintained by a small set of arbitrary videos. We optimize the two tasks with stochastic switching, using random reference-frame sampling and image-token dropout to prevent trivial first-frame copying. Our gradient analysis shows that the two objectives rapidly evolve toward nearly orthogonal update subspaces, explaining the stable optimization. Using CogVideoX-5B, we adapt a single model with 200K subject-image pairs and 4,000 arbitrary videos in 288 A100 GPU hours. This yields about 1% of compute compared to prior zero-shot baselines (i.e., 0.4% of VACE and 2.8% of Phantom) while using no subject-video pairs, yet remaining competitive in subject fidelity and motion quality. We show that the same recipe transfers to Wan 2.1-1.3B and Wan 2.2-5B.

Generating geometrically consistent videos remains an open challenge: text-to-video diffusion models trained on web-scale data treat geometry only implicitly, leading to object deformation, texture drift, and non-rigid backgrounds under camera motion. Existing solutions either improve consistency as a byproduct, apply only to static scenes or realign the latent space of the model completely. We introduce a geometry-consistency reward that directly measures whether motion in a generated video is compatible with a coherent scene. Our key insight is that in physically consistent videos, background motion should be explainable by rigid camera-induced flow, while independently moving objects should preserve appearance identity along motion trajectories. We operationalize this using optical flow, depth--pose predictions, and feature-based correspondence to separate rigid and dynamic regions and evaluate their respective consistency. Integrating this reward with reinforcement fine-tuning transforms geometric consistency from an emergent property into an explicit optimization objective for video generators. The approach is model agnostic and applies to diverse dynamic scenes containing both camera and object motion. Experiments show substantial reductions in temporal geometric artifacts over strong baselines while preserving perceptual quality. Code and model weights are published.

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.