Research map / Vision-language and multimodal / Multimodal LLMs
Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning
Shaofeng Yin, Jiaxin Ge, Zora Wang, Xiuyu Li, Chenyang Wang, Michael Black, Trevor Darrell, Angjoo Kanazawa, Haiwen Feng
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 |
|---|---|---|---|
| increase best-of-N search budget from 1 to 4 (ablation) | VIGA · BlenderBench Task 1 | VLM score: 1.44 -> 3.25 | Improved |
| add sliding-window multimodal memory (ablation) | VIGA · qualitative sequential generation | context retention | Improved |
| bound memory with a sliding window instead of unconstrained history (ablation) | VIGA · long-horizon generation | generation quality | Improved |
| use VIGA code-render-inspect loop instead of one-shot | Qwen3-8B · SlideBench | Execution: 0.10 -> 0.54 | Improved |
| use VIGA instead of one-shot | Claude-3.5 · SlideBench | Execution: 0.93 -> 0.90 | Hurt |
| use VIGA instead of BlenderAlchemy at best-of-1 | GPT-4o · BlenderBench Task 1 | VLM score: 1.75 -> 1.44 | Hurt |
| use VIGA instead of BlenderAlchemy at best-of-4 | GPT-4o · BlenderBench Task 2 | Photometric loss: 1.95 -> 2.94 | Hurt |
| use VIGA instead of one-shot | Qwen3-8B · BlenderGym Blend Shape | Photometric loss: 7.76 -> 13.51 | Hurt |
| use VIGA instead of one-shot | Gemini-2.5-Pro · BlenderGym Blend Shape | Photometric loss: 7.01 -> 7.62 | Hurt |
| use VIGA instead of one-shot | Gemini-2.5-Pro · BlenderBench Task 2 | Photometric loss: 10.14 -> 1.99 | Improved |
| use VIGA with Qwen3-VL-8B instead of BlenderAlchemy (best-of-1) | Qwen3-VL-8B · BlenderBench | average relative improvement: +312.20% | Improved |
| use VIGA instead of one-shot and BlenderAlchemy on BlenderGym | VIGA · BlenderGym | average improvement: +35.32% | Improved |
Closest papers by meaning
- BlenderGym: Benchmarking Foundational Model Systems for Graphics Editing (CVPR 2025 · similarity 0.86)
- MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.86)
- Generative Universal Verifier as Multimodal Meta-Reasoner (ICLR 2026 · similarity 0.86)
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.86)
- VIGC: Visual Instruction Generation and Correction (AAAI 2024 · similarity 0.86)
- Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.85)
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.
