History-Forcing & LightX2V: Open Source Long Video AI

History-Forcing (ziplab/History-Forcing) and LightX2V (ModelTC/LightX2V) tackle the two main barriers to long open source video production: temporal drift across 60 second rollouts and multi step inference latency. Developed by Zhejiang University and Microsoft Research (arXiv:2611.02782), History-Forcing introduces history contrast distillation (DMD on Wan2.1-T2V-1.3B and OPSD on DreamX-World-5B) so autoregressive video generators remember earlier chunks and track how a scene evolves out of sight when the camera pans away. Paired with the Apache 2.0 LightX2V inference engine, production teams can run 4 step NVFP4 and FP8 distilled video models at up to 50x speedup on a single RTX 5090 or H100 GPU.

Studio availability. History-Forcing and LightX2V are not available in AI FILMS Studio. This guide covers their open source releases and local deployment. In the Studio, use Seedance 2.0, Kling 3.0, Veo 3.1, Wan 2.6, or LTX-2 on the Text to Video page or in Studio Nodes.

1. History-Forcing: 60 Second Consistency & Scene Memory

Autoregressive video generators extend shots chunk by chunk, conditioning each new segment on its own previously generated frames. Over a 30 to 60 second rollout, small color shifts and geometric errors compound into severe drift, and objects that leave the camera frame often vanish or mutate when the camera looks back.

History-Forcing (xzxxntxdy/History-Forcing) fixes both failure modes during distillation.

• History contrast supervision. Evaluates the autoregressive teacher on the same noisy student chunk under two contexts (one with informative early history and one without) and adds the weighted residual to the distillation score or velocity target.
• 60 second consistent generation (dmd/model.pt). Built on Wan-AI/Wan2.1-T2V-1.3B, the DMD student anchors subject identity and lighting to early rollout chunks that have not suffered error accumulation.
• Scene evolution outside the frame (opsd/model.pt). Built on DreamX-World-5B (Wan2.2-TI2V-5B), static camera teacher rollouts simulate how moving actors or physical objects continue evolving while the student camera looks away, ensuring plausible world state upon revisit.

2. LightX2V: Up to 50x Faster Video Inference Across Wan, LTX-2, Hunyuan & H3

While History-Forcing solves temporal memory, LightX2V (ModelTC/LightX2V, Apache 2.0) provides the unified deployment engine and distilled checkpoints (huggingface.co/lightx2v) to run open source video models at interactive speeds.

• 50x speedup on RTX 5090 (Wan2.2-NVFP4-Sparse). Combines 4 step distillation without CFG with Blackwell NVFP4 quantization and sparse attention to accelerate Wan 2.2 14B (T2V and I2V) by over 50x compared to 40 step baseline inference.
• Autoregressive and world model profiles. Ships dedicated 2 step to 4 step scripts for Self-Forcing (FP8 and NVFP4), EdenLive, Matrix-Game-4, LingBot-Fast, StreamRender-H3, and TaoMate-H3.
• Support for major open backbones. Includes 4 step and 8 step Turbo LoRAs and quantized weights for MiniMax-H3 (768p audio and video), HunyuanVideo-1.5 (25x speedup plus LightTAE fast VAE), LTX-2 (FP8 plus block offload), and Qwen-Image-2.1 (including a native macOS app).

Checkpoints, Backbones & Hardware Matrix

Framework / Checkpoint	Hugging Face / GitHub ID	Base Model	Key Capability & Speed	Hardware & License
History-Forcing Long Video (DMD)	xzxxntxdy/History-Forcing	Wan2.1-T2V-1.3B	60 second consistent autoregressive video generation	Single 16 to 24 GB VRAM GPU · Open Research
History-Forcing World Model (OPSD)	xzxxntxdy/History-Forcing	DreamX-World-5B (Wan2.2-TI2V-5B)	Camera controlled world model with scene evolution	24 to 48 GB VRAM GPU · Open Research
LightX2V Engine & Distilled Hub	ModelTC/LightX2V / lightx2v	Wan 2.1/2.2, MiniMax H3, LTX-2, HunyuanVideo 1.5, Self-Forcing	4 step, FP8 / NVFP4 / GGUF, block offload, up to 50x speedup	RTX 4090 / RTX 5090 / H100 / Apple Silicon · Apache 2.0

Local Setup & 60 Second Generation Commands

To generate a 60 second video with History-Forcing (DMD), run the following commands.

git clone https://github.com/ziplab/History-Forcing.git
cd History-Forcing

# Download the Wan2.1-T2V-1.3B backbone and History-Forcing DMD checkpoint
hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir models/Wan2.1-T2V-1.3B
hf download xzxxntxdy/History-Forcing dmd/model.pt --local-dir models/History-Forcing

# Generate a 60 second consistent video rollout
bash scripts/long_video/inference.sh \
  --prompt "A musician strumming the strings of an acoustic guitar, lost in the melody of their song." \
  --output-dir output/demo/long_video

To deploy LightX2V via Docker or source for 4 step accelerated inference across Wan 2.2, LTX-2, HunyuanVideo 1.5, or MiniMax H3, run the following commands.

git clone https://github.com/ModelTC/LightX2V.git
cd LightX2V

# Example: Run 4 step distilled inference or launch the LightX2V service API
bash scripts/wan22/run_wan22_i2v_4step.sh

For deeper dives into interactive world models and lightweight consumer GPU video generation, read our guides to Helios real time long video generation, Tencent HunyuanVideo 1.5 on 14GB VRAM, FastVideo and FastWan step distillation, and ShotStream streaming multi shot video.

Long Form Sequences in AI FILMS Studio

History-Forcing and LightX2V run on local GPUs. To chain consistent shots with first and last frame locking, character references, and synchronized audio in the cloud, open the video workspace to run Seedance 2.0, Kling 3.0, Veo 3.1, Wan 2.6, and LTX-2, or build multi shot sequences in the Nodes Graph Editor.

---

Sources

• ziplab/History-Forcing GitHub Repository
• History Forcing Official Project Page
• History Forcing: Distilling History-Dependent Continuations for Long-Horizon Video World Models (arXiv:2611.02782)
• xzxxntxdy/History-Forcing on Hugging Face
• ModelTC/LightX2V GitHub Repository
• lightx2v Distilled & Quantized Models on Hugging Face