Research map / Image generation / Fast diffusion sampling
Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion
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 Diffusion Forcing, training a sequence diffusion model with independent per-token noise levels, instantiated as Causal Diffusion Forcing (RNN) for stable long-horizon video rollout, planning with Monte Carlo Guidance, and imitation learning.
Problem
Next-token prediction models are unstable on continuous data (e.g., video rollouts diverge) and cannot be guided toward objectives, while full-sequence diffusion models are non-causal and fixed-length, limiting variable-length generation and subsequence guidance.
Approach
Train a causal denoiser where each token gets an independent random noise level, so a single shared model can act as next-token predictor, full-sequence diffuser, or both; sampling uses per-token noise schedules (e.g., zig-zag) and enables guidance over future tokens and Monte Carlo Guidance.
Key results
- On D4RL Maze2D/Multi2D, Diffusion Forcing achieves 141.7 single-task and 146.2 multi-task average reward, above Diffuser (119.5 / 129.4).
- Diffusion Forcing achieves 80% success on the real Franka swap-fruit imitation task where diffusion policy fails, and 76% under occlusion/noise versus 48% for next-frame diffusion.
- Diffusion Forcing stably rolls out video up to about 1000 frames beyond training horizon while teacher-forcing and full-sequence diffusion baselines diverge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Monte Carlo Guidance (ablation) | Diffusion Forcing · Maze2D Single-task Average | reward: 129.67 -> 141.7 | Helped |
| add Monte Carlo Guidance (ablation) | Diffusion Forcing · Multi2D Multi-task Average | reward: 127.7 -> 146.2 | Helped |
| use Diffusion Forcing instead of Diffuser* | offline planning · Maze2D U-Maze | reward: 113.9 \(\pm\) 3.1 -> 116.7 \(\pm\) 2.0 | Helped |
| use Diffusion Forcing instead of Diffuser* | offline planning · Maze2D Medium | reward: 121.5 \(\pm\) 2.7 -> 149.4 \(\pm\) 7.5 | Helped |
| use Diffusion Forcing instead of Diffuser* | offline planning · Maze2D Large | reward: 123.0 \(\pm\) 6.4 -> 159.0 \(\pm\) 2.7 | Helped |
| use Diffusion Forcing instead of Diffuser* | offline planning · Multi2D U-Maze | reward: 128.9 \(\pm\) 1.8 -> 119.1 \(\pm\) 4.0 | Hurt |
| use Diffusion Forcing instead of Diffuser* | offline planning · Multi2D Medium | reward: 127.2 \(\pm\) 3.4 -> 152.3 \(\pm\) 9.9 | Helped |
| use Diffusion Forcing instead of Diffuser* | offline planning · Multi2D Large | reward: 132.1 \(\pm\) 5.8 -> 167.1 \(\pm\)2.7 | Helped |
| use Diffusion Forcing for imitation learning with memory | robot fruit swapping · Robot fruit swapping | success rate | Helped |
| mark perturbed or missing observations as noisy | Diffusion Forcing · Robot fruit swapping with noisy or missing observations | success rate: 76\% vs. 48\% | Helped |
Surprising findings
- Diffuser's performance collapses (e.g., 8.7 single-task average) when executing its generated actions directly, requiring a hand-crafted PD controller.
Limitations
- Causal implementation is based on a vanilla RNN; transformer scaling not investigated.
- No investigation of scaling to internet-scale datasets or higher-resolution video.
- Diffusion-based video tasks evaluated on relatively small domains (Minecraft, DMLab).
Experimental setup
Datasets: Minecraft gameplay videos DMLab navigation videos D4RL Maze2D Multi2D Franka robot teleoperation dataset multivariate time series forecasting benchmark
Benchmarks: D4RL Maze2D Multi2D Minecraft video rollout DMLab video rollout real robot imitation task multivariate time series forecasting
Metrics: average normalized reward success rate rollout stability
Baselines: MPPI CQL IQL Diffuser next-frame diffusion (teacher forcing) full-sequence diffusion diffusion policy
Builds on: DDPM Diffuser teacher forcing classifier guidance diffusion policy
Closest papers by meaning
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.88)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.88)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.88)
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.88)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.88)
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.