scope-causal-forcing

A Daydream Scope plugin that adds the Causal Forcing pipeline for real-time streaming video generation.

Real-time streaming generation with Causal Forcing on Wan2.1-1.3B (17 FPS on H100, ~10 FPS on RTX 4090). Video from the Causal Forcing project page.

What is Causal Forcing?

Causal Forcing (Tsinghua / Shengshu / UT Austin, Feb 2026) is the successor to Self-Forcing -- the training method behind Scope's built-in LongLive pipeline. It fixes a theoretical flaw in Self-Forcing's ODE initialization by using an autoregressive teacher instead of a bidirectional one, producing strictly better video quality at identical inference speed.

Improvements over Self-Forcing (same GPU, same FPS):

Metric	Improvement
Dynamic Degree	+19.3%
VisionReward	+8.7%
Instruction Following	+16.7%

Both models run on Wan2.1-T2V-1.3B with 4-step denoising and KV-cached autoregressive generation.

Requirements

• NVIDIA GPU with 20+ GB VRAM (RTX 3090 / 4090 / 5090 or cloud equivalent)
• Daydream Scope installed

Installation

uv pip install git+https://github.com/daydreamlive/scope-causal-forcing.git

Or for local development:

git clone https://github.com/daydreamlive/scope-causal-forcing.git
cd scope-causal-forcing
uv pip install -e .

The plugin registers automatically via entry points — restart Scope and causal-forcing will appear in the pipeline selector.

Model Weights

On first load, Scope will download the required model weights:

Model	Source	Size
Wan2.1-T2V-1.3B	Wan-AI/Wan2.1-T2V-1.3B	~3 GB
UMT5-XXL encoder	google/umt5-xxl	~10 GB
Causal Forcing checkpoint	zhuhz22/Causal-Forcing	~3 GB
Wan2.1 VAE	(bundled with Wan2.1-T2V-1.3B)	~300 MB

If you already use LongLive in Scope, the base Wan2.1 and UMT5 weights are shared — only the Causal Forcing checkpoint is an additional download.

Configuration

Parameter	Default	Description
height	480	Output height in pixels
width	832	Output width in pixels
denoising_steps	[1000, 750, 500, 250]	4-step warped denoising schedule
vae_type	wan	Full VAE (wan) or 75% pruned (lightvae)
base_seed	42	Random seed for reproducibility

How It Works

Each frame is generated autoregressively with KV caching:

1. Denoise — 4-step spatial denoising loop (flow matching with warped timesteps)
2. Cache — Re-run at timestep=0 to write clean context into KV cache
3. Decode — Wan VAE decodes latent to pixels with temporal caching
4. Advance — Move to next frame

This is the same inference architecture as LongLive/Self-Forcing — only the weights differ.

References

• Paper: Causal Forcing: Autoregressive Distillation via Causal ODE
• Code: github.com/thu-ml/Causal-Forcing
• Weights: huggingface.co/zhuhz22/Causal-Forcing

License

Apache-2.0 (same as the Causal Forcing model weights)