Research map / Video / Video generation
EvoID: Reinforced Evolution for Identity-Preserving Video Generation
Yiheng Zhang, Zhaofan Qiu, Zunxu Liu, Yingwei Pan, Ting Yao, Tao Mei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RL with uniform-weighted objective rewards (ablation) | VACE-14B · OpenS2V-Eval Human-Domain | Total Score / EvoScore: 0.658 -> 0.685; 0.544 -> 0.627 | Improved |
| add frozen T2V teacher regularization to uniform-weighted rewards (ablation) | EvoID objective-reward variant · OpenS2V-Eval Human-Domain | EvoScore: 0.627 -> 0.663 | Improved |
| replace uniform with prior-weighted rewards (with teacher) (ablation) | EvoID objective-reward variant · OpenS2V-Eval Human-Domain | Total Score: 0.706 -> 0.715 | Improved |
| add teacher prior to prior-weighted rewards (ablation) | EvoID objective-reward variant · OpenS2V-Eval Human-Domain | Total Score: 0.715 -> 0.715 (EvoScore 0.645 -> 0.664) | No clear effect |
| add MLLM-based preference reward (VQScore) (ablation) | EvoID with objective rewards · OpenS2V-Eval Human-Domain | Total Score / EvoScore: Total 0.715 -> 0.685; EvoScore 0.664 -> 0.675 | Mixed |
| add MLLM-directed dynamic weighting (ablation) | EvoID with OR + PR · OpenS2V-Eval Human-Domain | Total Score / EvoScore: 0.685 -> 0.704; 0.675 -> 0.682 | Improved |
| add EvoScore as a training reward (ablation) | EvoID · OpenS2V-Eval Human-Domain | EvoScore / Total Score: EvoScore 0.682 -> 0.718; Total 0.704 -> 0.687 | Mixed |
| add EvoScore as a training reward (ablation) | EvoID · OpenS2V-Eval Human-Domain | Face Similarity: 0.745 -> 0.675 | Hurt |
| use EvoID RL training instead of base VACE-14B | VACE-14B · OpenS2V-Eval Human-Domain | Motion Amplitude: 0.118 -> 0.350 | Improved |
| use EvoID instead of ViduQ2 in human evaluation | EvoID · Human 2AFC | video quality win rate: 45% | No clear effect |
Closest papers by meaning
- SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents (NeurIPS 2026 · similarity 0.88)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.85)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.85)
- EvoGround: Self-Evolving Video Agents for Video Temporal Grounding (NeurIPS 2026 · similarity 0.85)
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (NeurIPS 2026 · similarity 0.85)
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
