Video World Models · 2026-10-12
Golden Anchor for Video Generation
Jinyuan Xu, Shitong Shao, Lichen Bai, Zikai Zhou, Guoxi Huang, Nantheera Anantrasirichai, Zeke Xie
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study how “anchor” images (produced by a text-to-image model) affect image-to-video (I2V) generation, and find that image aesthetic/quality scores weakly correlate with downstream video quality—especially structure/texture, which is negatively correlated. They propose Golden Anchor Learning (GAL), which fine-tunes a T2I generator using downstream video rewards via offline preference learning (DPO). GAL improves VBench and video/frame reward scores and yields consistent motion gains.
Why it matters
Many controllable video pipelines rely on choosing or generating an anchor image using image-domain metrics, which the authors show can misalign with downstream video quality and motion. GAL changes the upstream anchor distribution using video-derived feedback while keeping the I2V model frozen, improving both video and motion quality across multiple I2V models (including transfer to unseen renderers).
Method
- Identify misalignment: across 49,416 prompt–anchor–video records, UniPercept aesthetics/quality are nearly uncorrelated with VideoReward, while UniPercept structure-and-texture is negatively correlated with both visual and motion quality.
- Golden Anchor Learning (GAL): sample anchor candidates from a T2I generator, render each once with a frozen I2V model, score videos with a frozen video-reward model, and convert video scores into image-level preference pairs for offline DPO.
- Train only the upstream T2I (video feedback via DPO), keeping the I2V renderer frozen; the default signal uses VideoReward-Overall for preference construction.
Limitation
The authors report that they train only on HunyuanVideo-1.5 renders, and they extend gains to other I2V models; however, detailed training/labeling assumptions beyond this single label source are not specified in the provided text.
Abstract (from arXiv)
Video generation supports digital content creation, with applications in virtual avatars, AI-driven filmmaking, interactive game engines, and world models. State-of-the-art systems achieve controllability primarily through image-to-video (I2V) generation, in which an injected anchor image determines the identity, composition, and style of the resulting video. In modern pipelines, this anchor is synthesized by a text-to-image (T2I) generator. This raises a largely overlooked question: Are images obtained directly from the T2I model truly suitable for current video models? Our analysis shows that anchor aesthetic and quality scores are weakly correlated with downstream video quality. The structure-and-texture score is negatively correlated with I2V video quality in both the visual and motion dimensions. Fine-tuning the T2I model on an image reward accordingly yields anchors with visibly denser structural and character detail: they score higher on image metrics, yet lower the video reward on every I2V model we test and change downstream motion unpredictably. We therefore introduce Golden Anchor Learning (GAL), to our knowledge the first framework for learning golden anchors by optimizing a T2I generator with downstream video feedback. GAL evaluates candidate anchors through the videos they induce under a frozen I2V model and converts video rewards into image-level preferences for DPO fine-tuning of the T2I generator. On all four I2V models, GAL raises the VBench average by $2.97\%$--$8.46\%$ over the untuned generator. The rendered clips also score higher under VideoReward-Overall, and their extracted frames under HPSv3 and UniPercept. GAL yields consistent motion gains across all four I2V models, as measured by a $5.62\%$--$38.24\%$ increase in VBench dynamic degree. Although GAL is trained only on HunyuanVideo-1.5 renders, these gains extend to the other three I2V models.
Related papers
- LongTake: Learning to Sustain Dynamics in Long-Horizon Video Generation
- ACWM-RM: A Reference-Free Reward Model from Complementary Mismatch Preferences
- Bootstrapping Video Interaction Generation with Synthetic State Transitions
- Video-Coder: Reinforcing Agentic Coding for Video Generation
- FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-09 · Indexed by Pub-AI: 2026-10-12 UTC
Golden Anchor for Video Generation
Preprint posted: 2026-10-09 · Original record ↗
- arxiv
- 2610.13217
Explicit source versions (1)
- v2 · 2026-10-12
- Preprint available
Source date: 2026-10-09 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
- Explicit source version detected · v2
Source date: 2026-10-12 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
Abstract · arxiv
Video generation supports digital content creation, with applications in virtual avatars, AI-driven filmmaking, interactive game engines, and world models. State-of-the-art systems achieve controllability primarily through image-to-video (I2V) generation, in which an injected anchor image determines the identity, composition, and style of the resulting video. In modern pipelines, this anchor is synthesized by a text-to-image (T2I) generator. This raises a largely overlooked question: Are images obtained directly from the T2I model truly suitable for current video models? Our analysis shows that anchor aesthetic and quality scores are weakly correlated with downstream video quality. The structure-and-texture score is negatively correlated with I2V video quality in both the visual and motion dimensions. Fine-tuning the T2I model on an image reward accordingly yields anchors with visibly denser structural and character detail: they score higher on image metrics, yet lower the video reward on every I2V model we test and change downstream motion unpredictably. We therefore introduce Golden Anchor Learning (GAL), to our knowledge the first framework for learning golden anchors by optimizing a T2I generator with downstream video feedback. GAL evaluates candidate anchors through the videos they induce under a frozen I2V model and converts video rewards into image-level preferences for DPO fine-tuning of the T2I generator. On all four I2V models, GAL raises the VBench average by $2.97\%$--$8.46\%$ over the untuned generator. The rendered clips also score higher under VideoReward-Overall, and their extracted frames under HPSv3 and UniPercept. GAL yields consistent motion gains across all four I2V models, as measured by a $5.62\%$--$38.24\%$ increase in VBench dynamic degree. Although GAL is trained only on HunyuanVideo-1.5 renders, these gains extend to the other three I2V models.
