Research map / 3D vision / 3D generation

PhotoFlow: Agentic 3D Virtual Photography Missions

NeurIPS 2026 · 3D vision › 3D generation · approach: Generative indoor scene synthesis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces PhotoFlow, a Director-Reviewer-Reflector LLM/VLM agent that searches executable Blender camera states for language-conditioned virtual photography, plus the VPhotoBench benchmark (47 scenes, 141 missions).

Problem

Language-conditioned virtual photography requires an agent to pick an executable camera state in an arbitrary 3D scene that satisfies spatial constraints and abstract aesthetic intent, a combination not covered by existing benchmarks.

Approach

A closed-loop agent with a Director that proposes candidate cameras from scene scouting, soft blueprints and anchor banks; a Reviewer combining projection rule checks, VLM visual critique, and pairwise incumbent selection; and a Reflector that maintains region memory, suppresses dead zones, and forces high-explore relocation under a fixed rendering budget.

Key results

• PhotoFlow achieves M_qs 0.578 and Succ@0.55 0.622 on 90 common completed held-out tasks, versus 0.567/0.567 for single-chain reflection and 0.514/0.400 for single-step LLM.
• In human selection, PhotoFlow receives 31.33% aesthetic selections versus 24.89% for the strongest non-ours baseline (single-chain).
• Removing region memory lowers M_qs to 0.572 and Succ@0.55 to 0.544, and raises revisit rate to 0.644.

Surprising findings

• Disabling high-explore relocation slightly raises mean M_qs (0.584) and Succ is lower only modestly, while substantially reducing coverage and increasing collapse, so the component is a safeguard rather than a universal gain.
• Stricter heuristic early stopping gave no reliable quality benefit during development.

Limitations

• Global exploration is bounded by the anchor bank quality from scene scouting.
• Main results are on a common-completed subset of 90 of 141 tasks because 27 tasks failed to render, and no end-to-end availability score is reported.
• Internal Reviewer scores are not sufficient evidence; no formal significance tests or threshold sensitivity analyses; virtual scenes may not transfer to physical robots or dynamic scenes.

Experimental setup

Datasets: Blender Demo Files archive Blend Swap

Benchmarks: VPhotoBench

Metrics: M_qs quality-alignment composite Succ@0.55 IAA IQA ISTA coverage collapse revisit

Baselines: Single-Step LLM Iterative Single-Chain Reflection Anchor Bank Best-of-N Random Search

Builds on: NavGPT UniPercept

Closest papers by meaning

• IR3D-Bench: Evaluating Vision-Language Model Scene Understanding as Agentic Inverse Rendering (NeurIPS 2025 · similarity 0.86)
• SceneCraft: An LLM Agent for Synthesizing 3D Scenes as Blender Code (ICML 2024 · similarity 0.85)
• 3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code (NeurIPS 2026 · similarity 0.85)
• SceneWeaver: All-in-One 3D Scene Synthesis with an Extensible and Self-Reflective Agent (NeurIPS 2025 · similarity 0.84)
• Muses: 3D-Controllable Image Generation via Multi-Modal Agent Collaboration (AAAI 2025 · similarity 0.84)
• VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis (CVPR 2026 · similarity 0.84)