Research map / Video / Video generation
Reasoning Diffusion for Unpaired Test Time Out-of-distribution Text-Image to Video Generation
Zirui Pan, Xin Wang, Yipeng Zhang, Hong Chen, Kecheng Zheng, Wenwu Zhu
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 VisionNarrator, AlignFormer and two-stage training to Wan2.1 | Wan2.1-I2V-14B · ActivityNet (unpaired inputs) | CLIP Score (Text): 0.224 -> 0.261 | Improved |
| add VisionNarrator, AlignFormer and two-stage training to Wan2.1 | Wan2.1-I2V-14B · ActivityNet (unpaired inputs) | User Rank: 2.692 -> 1.743 | Improved |
| add ReasonDiff modules to Wan2.1 on paired inputs | Wan2.1-I2V-14B · MSR-VTT (paired inputs) | Imaging Quality: 0.560 -> 0.571 | Improved |
| add ReasonDiff modules to Wan2.1 on image similarity | Wan2.1-I2V-14B · ActivityNet (unpaired inputs) | CLIP Score (Image): 0.518 -> 0.528 | No clear effect |
| add the auxiliary latent reconstruction loss (second training stage) (ablation) | ReasonDiff · ActivityNet (unpaired inputs) | all VBench and CLIP metrics (esp. Motion Smooth) | Improved |
| add per-frame narratives from VisionNarrator instead of single user prompt (ablation) | ReasonDiff · ActivityNet (unpaired inputs) | Dynamic Degree | Improved |
| add reasoning-enhanced latents from AlignFormer on top of narratives (ablation) | ReasonDiff · ActivityNet (unpaired inputs) | Imaging Quality | Improved |
| rewrite the text prompt with an MLLM instead of using VisionNarrator and AlignFormer (ablation) | Wan2.1 base model · ActivityNet (unpaired inputs) | CLIP Score (Text) | Hurt |
| rewrite prompt and manually choose condition index for Wan2.1 (naive approach) (ablation) | Wan2.1 · qualitative unpaired examples | qualitative coherence | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Dynamicrafter — DynamiCrafter: Animating Open-domain Images with Video Diffusion Priors (ECCV 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- OpenCoF: Learning to Reason Through Video Generation (NeurIPS 2026 · similarity 0.85)
- Hierarchical Spatio-temporal Decoupling for Text-to-Video Generation (CVPR 2024 · similarity 0.85)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.85)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.85)
- Generative Universal Verifier as Multimodal Meta-Reasoner (ICLR 2026 · similarity 0.84)
- DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · 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.
