AsadIsmail/Wan2.2-TI2V-5B-ternary

Wan2.2-TI2V-5B — Ternary Quantized (tritplane3)

Packed ternary version available. The packed/ folder now contains the real 2-bit packed transformer (1.78x smaller: 5.5 GB vs 9.8 GB FP16) that runs on the ternary-quant packed runtime. Load it with packed/load_packed_dit.py (pip install ternary-quant). The root files remain FP16 so stock diffusers still works. Quality note: the packed transformer was quantized without activation calibration (memory-minimal streaming), so treat it as experimental.

First publicly available ternary-quantized Wan 2.2 model on HuggingFace.

Ternary-quantized version of Wan-AI/Wan2.2-TI2V-5B-Diffusers — Alibaba's latest text-image-to-video DiT model (5B, 572 likes on original).

Specifications

Property	Value
Base Model	Wan-AI/Wan2.2-TI2V-5B-Diffusers
Architecture	WanTransformer3DModel (DiT)
Transformer Params	5.00B
Quantization	tritplane3 (306 linear layers)
Text Encoder (UMT5-XXL)	FP16 (preserved)
VAE (WanVAE)	FP16 (preserved)
License	Apache 2.0

Size

Method	Transformer Size
FP16 (original)	10.02 GB
Ternary tritplane3 (theoretical packed)	~5.0 GB
In this repo (dequantized FP16)	9.4 GB

Usage

import torch
from diffusers import WanPipeline
from diffusers.utils import export_to_video

pipe = WanPipeline.from_pretrained(
    "AsadIsmail/Wan2.2-TI2V-5B-ternary",
    torch_dtype=torch.bfloat16,
)
pipe.to("mps")  # or "cuda"

output = pipe(
    prompt="a cat walking on green grass",
    num_frames=81,
    num_inference_steps=30,
).frames[0]
export_to_video(output, "output.mp4", fps=16)

Collection

Part of ternary-models.