Penn's open-source HARMONY rebuilds 3D scenes from one image, matching GPT-6 Astra
thoma_gu · x · 2026-09-24
Researchers at the University of Pennsylvania released HARMONY, an open-source framework that reconstructs compositional 3D scenes from a single indoor image, plus the HARMONY300 benchmark of 300 scenes.
- Hierarchical agentic reasoning: a VLM agent builds scenes like a 3D artist — calibrating the camera, recovering room layout, then placing wall-mounted objects, freestanding furniture, and dependent decorations in order
- Self-reflection loop: after each placement the agent renders and compares against the input image, reflects on errors, and iteratively corrects to prevent error accumulation; depth-first traversal is used for furniture
- Results: even with older VLM backbones, it matches GPT-6 Astra in visual quality while achieving better geometric alignment
Paper, code, data, and benchmark are all open-sourced.
Related event: UPenn Open-Sources HARMONY for Single-Image 3D Scene Reconstruction(3 posts)→
More from Multimodal
- Side project ports most video generation models from PyTorch to Jax for TPU — ceciletamura · 2026-09-24
- Qwen Image 2.1 masked inpainting: working crop-and-stitch graph, wiring and prompting differences from Flux — Reasonable_Arm7239 · 2026-09-24
- WIP H3 Max spatial reframing tool's artifacts create painterly textures — adamho · 2026-09-24
- Counterintuitive: generic synthetic data beats self-generated data for distillation adapters — ostrisai · 2026-09-24
- Recraft V4.1 flash image generation wows with quality — aziz4ai · 2026-09-24
- UCLA x Google Cloud AI Launch PaperBanana-Interact for Chat-Based Paper Diagram Refinement — VioletNPeng · 2026-09-24