ar-tf-step9500

AR-TF / causal-DN step 9500

This repository contains the exported step-9500 checkpoint from the Stage-1 causal autoregressive DN teacher-forcing experiment used in the EasyLab AR-TF research line.

The short repository name is ar-tf-step9500; the exact experiment is the causal-DN teacher-forcing model tagged 9src1k--step16500-init--chunk3--480x832--dnfix. This checkpoint also served as the initialization for later AR-TF, offline, DMD, OPSD, Warp-as-History, and correspondence experiments.

Code and documentation

• Source branch: YuryLee/easy-lab-hzh-video_gen — release/ar-tf-step9500
• Full handoff guide: docs/ar_tf_step9500_handoff.md
• Base model: Wan-AI/Wan2.2-TI2V-5B-Diffusers

Files

SHA-256 for model.pt:

27520c4892e867e07f1e6c81f36fb71be2a817cc45e79838c83ef7db14631d8e

model.pt is approximately 22.9 GiB. It uses the PyTorch pickle container; only load a trusted download and verify SHA256SUMS first.

Reference recipe

• Architecture base: Wan2.2 TI2V-5B Diffusers
• Resolution: 480 x 832
• Clip length: 81 frames
• Autoregressive chunk size: 3 latent frames
• Objective: causal DN teacher forcing with chunk-local DN cross-attention
• Warm start: bidirectional DN step 16500
• Data mode: pre-encoded latent caches
• Mixture: nine sources, approximately 1,000 clips per source
• Training backend: native PyTorch FSDP2
• Reference hardware: 8 x A800 80 GB
• Precision: bfloat16

Download

huggingface-cli download DraCary0625/ar-tf-step9500 \
  --local-dir checkpoints/ar-tf-step9500

sha256sum -c checkpoints/ar-tf-step9500/SHA256SUMS

This is a custom EasyLab checkpoint, not a standalone Diffusers pipeline. It cannot be loaded directly with DiffusionPipeline.from_pretrained().

Inference

Clone the matching code branch and install the research dependencies:

git clone --branch release/ar-tf-step9500 \
  https://github.com/YuryLee/easy-lab-hzh-video_gen.git
cd easy-lab-hzh-video_gen
pip install -r requirements-research.txt
pip install -e .

The released inference path consumes EasyLab Parquet latent caches containing pre-encoded text embeddings, DN latents, and first-frame latents:

CUDA_VISIBLE_DEVICES=0 python -u scripts/inference_causal_ar.py \
  /path/to/checkpoints/ar-tf-step9500 \
  outputs/ar_tf_9500_demo \
  --sources /absolute/path/to/a/latent-cache \
  --indices 0 \
  --steps 50 \
  --guidance 5.0 \
  --device cuda:0

See the handoff guide for the tested package versions, base-model download, storage-path mapping, multi-GPU commands, training config lineage, and checkpoint-loading details.

Intended use

The checkpoint is intended for research on autoregressive video generation, long-horizon causal sampling, memory mechanisms, and post-training methods. It is not presented as a production service or a safety-filtered end-user model.

Limitations

• A compatible Wan2.2 base-model directory is required.
• The default released inference path requires precomputed latent-cache input.
• Single-process inference was observed to require roughly 40-45 GB of GPU memory.
• Long-video generation is slow because chunks are sampled autoregressively.
• The full training recipes target multi-GPU FSDP2 systems and depend on datasets and intermediate checkpoints that are not included here.

License and data

The EasyLab code and referenced Wan2.2 base model are distributed under the Apache-2.0 license. Dataset assets and source videos are not redistributed by this repository and remain subject to their original licenses and terms.