Percentile-Calibrated HiFloat4 W4A4 PTQ for Wan2.2 Text-to-Video Generation
Official quantized checkpoint release for our award-winning solution to the ICME 2026 Low-Bit-width Large-Model Quantization Challenge:
Percentile-Calibrated HiFloat4 W4A4 Post-Training Quantization for Wan2.2 Text-to-Video Generation
Overview
This repository contains the official minimal PTQ checkpoint for the public release of the challenge solution above.
-
Code repository: https://github.com/GuoShuaiGO/percentile-hif4-wan22-icme2026
-
Checkpoint repository: https://huggingface.co/GuoShuaiGO/percentile-hif4-w4a4-wan22
-
Base model: Wan-AI/Wan2.2-I2V-A14B-Diffusers
-
Quantized artifact: quant_ckpt.pth
-
Release config: hif4_w4a4.yaml
The checkpoint stores only the inference-time PTQ state needed to reconstruct the released quantized model on top of the public Wan2.2 base model. It does not duplicate the full Wan base weights.
Files
quant_ckpt.pth: minimal inference-time PTQ checkpointhif4_w4a4.yaml: quantization configuration used for the released setting
Companion Code Release
Use this checkpoint with the companion official code repository for this release.
The code repository provides:
- the quantized inference entry point
- the vendored HiFloat4 runtime source
- build scripts for the required custom operators
- public demo prompts and reproduction instructions
Companion code repository: https://github.com/GuoShuaiGO/percentile-hif4-wan22-icme2026
Public Reproducibility Note
The competition task is text-to-video generation, while the public base checkpoint used in this release is Wan2.2-I2V-A14B-Diffusers, whose public pipeline requires an image input. The companion code reproduces prompt-only public inference by feeding an explicit blank placeholder image and logging that behavior during generation.
Example Usage
After downloading the public Wan2.2 base model and cloning the companion code repository:
1mkdir -p checkpoints
2huggingface-cli download GuoShuaiGO/percentile-hif4-w4a4-wan22 quant_ckpt.pth --local-dir ./checkpoints
3huggingface-cli download GuoShuaiGO/percentile-hif4-w4a4-wan22 hif4_w4a4.yaml --local-dir ./checkpoints
4
5python examples/wan2.2_hif4/quant_inference.py \
6 --model-path ./models/Wan2.2-I2V-A14B-Diffusers \
7 --quant-ckpt ./checkpoints/quant_ckpt.pth \
8 --dataset-json ./assets/demo_prompts.json \
9 --prompt-key cap \
10 --num-samples -1 \
11 --height 720 \
12 --width 1280 \
13 --num-frames 61 \
14 --num-inference-steps 40 \
15 --output-dir ./logs/wan2.2_hif4/generated_videos
License
This checkpoint release is distributed under the Apache-2.0 License.
