Self Gradient Forcing¶

Development line: project:self-gradient-forcing · thread self-gradient-forcing-development
Last event: 2026-07-24 · 1 dated since 2026-07-24 · Researched: 2026-09-05 · confidence: high

What it is¶

Self Gradient Forcing (SGF) is a training method for autoregressive video diffusion models. It replays sampled self-generated context so future-frame losses train causal KV memory without retaining gradients through the full rollout.

• Code and checkpoints: built on Wan2.1 T2V 1.3B and 14B foundations in framewise and chunkwise settings.
• Horizon: 5 seconds reported training window, with demonstrations and evaluation extending to 60 and 240 seconds.

Use it as a research implementation for long-horizon video extrapolation, not as a drop-in hosted video generator.

Development line¶

• 2026-07-24 — Self Gradient Forcing project resources were linked. On 2026-07-24, the project page, GitHub repository, and Hugging Face model page were linked. They provide the project overview, code, and model weights in one place. The links do not show whether this was an initial release, an update, or another technical milestone.

What changed¶

• 2026-07-22 — the paper introduced bounded two-pass reconstruction to bridge the historical context-gradient gap in frozen-cache Self Forcing.
• 2026-07-23 — the authors publicly released the paper, inference and training code, and model checkpoints.
• 2026-07-24 — the linked project page, repository, and model collection made the release discoverable, specifying framewise and chunkwise releases rather than one generic checkpoint.

How to use this¶

As of 2026-07-24, practitioners can use the linked project page, GitHub repository, and Hugging Face page to locate Self Gradient Forcing's public overview, implementation, and model resources.

1. Create a Python 3.10 environment, install the repository requirements and FlashAttention, then install the project in development mode. — https://github.com/zhuang2002/Self_Gradient_Forcing
2. Run scripts/download_weights.sh; it retrieves Wan2.1 T2V 1.3B/14B bases, Causal-Forcing initializations, SGF checkpoints, and prompts. — https://github.com/zhuang2002/Self_Gradient_Forcing
3. Choose the matching release mode and run bash scripts/infer_self_gradient_forcing.sh framewise or chunkwise; use the matching configuration and checkpoint. — https://github.com/zhuang2002/Self_Gradient_Forcing
4. For training or adaptation, launch the corresponding framewise or chunkwise training script and preserve the rollout/reconstruction split defined by SGF. — https://github.com/zhuang2002/Self_Gradient_Forcing

Best practices¶

• Evaluate long-horizon behavior separately from short clips: the paper treats 5-second VBench as a sanity check and focuses the main comparison on 60- and 240-second extrapolation, where memory errors accumulate. — https://arxiv.org/abs/2607.20368
• Keep the release setting, configuration, and checkpoint aligned: framewise and chunkwise variants use different cache geometries and are not interchangeable. — https://github.com/zhuang2002/Self_Gradient_Forcing
• Use EMA weights for released or trained checkpoints unless deliberately comparing raw generator weights. — https://github.com/zhuang2002/Self_Gradient_Forcing

Superseded by this¶

• 2026-07-22 — frozen-cache Self Forcing is superseded by SGF only for the missing context-gradient path: SGF preserves the no-gradient serial rollout but reconstructs the sampled computation so future losses train memory writing. This is not a general deprecation of all Self Forcing workflows.

Still unknown¶

• Independent results: no outside reproduction covers quality, compute cost, or compatibility; reported long-horizon gains remain author-reported.
• Production support: the release material does not establish a stable production support policy or a maintained inference service.

Sources¶

Agent brief¶

• Subject: project:self-gradient-forcing, thread self-gradient-forcing-development, 1 dated events 2026-07-24 → 2026-07-24.
• Practical note: As of 2026-07-24, practitioners could use the linked project page, GitHub repository, and Hugging Face page to locate Self Gradient Forcing's public overview, implementation, and model resource.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.