• The ImageNet-FID Leaderboard May Point the Wrong Way. DiffusionBench trained 21 diffusion models in one framework and found ImageNet rank and text-to-image rank aren't weakly correlated — they're negatively correlated, with Pearson coefficients as low as -0.58. Stop picking generation methods on a single benchmark.

• A 3D Scene From One Photo Now Drops Straight Into a Game Engine. FLAT decodes the geometry in a video-diffusion latent into surfaced triangle meshes in a single feedforward pass. A light refinement step turns the output into a real-time renderable asset that plugs into a standard graphics pipeline.

• More Photorealistic Doesn't Mean More Understanding. CF-World's counterfactual benchmark shows every model collapses once the rules break common sense. Visual realism comes from pattern matching, not causal reasoning.

Also Notable

• Cross-Chart RAG Finally Gets a Real Benchmark. Existing benchmarks either test only structured tables or stitch questions from extracted key points, leaving queries and evidence overlapping at the word level with incoherent reasoning chains. ChartWalker uses a layered knowledge graph to build questions that genuinely need cross-chart reasoning, with code.
• Diffusion Alignment Restores High-Frequency Detail in Image-to-3D Gaussians. Sparse-voxel methods are limited by discriminative 2D features and tend to lose the input image's high-frequency detail. FLUX3D switches to a diffusion-aligned sparse representation for fidelity, a second technical route for 3D generation alongside today's FLAT.
• A Generative Model Fills the Missing Frames in Cyclone Monitoring. Microwave satellite imagery has long revisit gaps and easily misses a cyclone's rapid-intensification window. MotifGen does spatiotemporal interpolation across multiple sources to fill the gaps, a concrete deployment from INRIA.

Read the full edition →