Research map / Image generation / Autoregressive and latent generation

Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Diffusion planning and policies · 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.

Proposes CoFi, an inference-time compositional diffusion sampler that first aligns local denoised estimates to a shared coarse scaffold and then re-noises and refines it with the same local prior, improving global coherence at 2–8x fewer denoiser evaluations than CDGS.

Problem

Compositional diffusion from short-horizon priors yields locally consistent overlaps but often globally implausible long-horizon outputs, and existing fixes (repeated inner message passing or inference-time optimization) become expensive as the number or dimensionality of local plans grows.

Approach

Replace repeated local consistency propagation with a closed-form shared-scaffold alignment of local denoised estimates (shrinking each toward the common mean), then partially re-noise the scaffold to an intermediate step t* and denoise it with the same pretrained local prior to restore fine detail.

Key results

• On OGBench PointMaze-Giant, CoFi reaches 96% success vs 82% for CDGS; Scene-Play improves from 51% to 63%.
• On panoramas, Intra-LPIPS drops from 0.65 (CDGS) to 0.48 and CLIP score rises from 32.67 to 33.18.
• On 273-frame video, subject consistency improves from 91.67 (CDGS) to 94.11 while using 2–8x fewer denoiser evaluations than CDGS.

Surprising findings

• On AntMaze-Giant and PointMaze-Medium/Large, CoFi is essentially on par with CDGS (saturated or comparable), so gains concentrate on the hardest settings.

Limitations

• Output quality is bounded by the expressiveness of the pretrained local prior.
• Consensus schedule lambda_t and refinement depth t* need per-domain tuning.
• Uniform weighting of local plans may be suboptimal when segments differ in structural importance.

Experimental setup

Datasets: OGBench VBench prompts

Benchmarks: OGBench VBench

Metrics: success rate Intra-LPIPS Intra-Style loss CLIP score subject consistency temporal flickering aesthetic quality prompt alignment

Baselines: GCBC GCIVL GCIQL HIQL GSC CD CDGS MultiDiffusion SyncDiffusion Gen-L-Video CogVideoX-2B

Builds on: CDGS GSC DDIM Bethe approximation SDEdit

Closest papers by meaning

• Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.88)
• Compositional Monte Carlo Tree Diffusion for Extendable Planning (NeurIPS 2025 · similarity 0.88)
• Predictive but Not Plannable: RC-aux for Latent World Models (NeurIPS 2026 · similarity 0.87)
• DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026 · similarity 0.87)
• Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.87)
• LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE (NeurIPS 2026 · similarity 0.87)