Wan2.2-T2V-A14B-BF16

Wan2.2-T2V-A14B — BF16 Variant

This is a BF16-precision variant of Wan-AI/Wan2.2-T2V-A14B-Diffusers, created to reduce storage and VRAM requirements while maintaining full inference quality on modern hardware.

What changed

Total transformer storage drops from ~114 GB to ~58 GB. All other files (configs, tokenizer, scheduler) are identical to the original.

BF16 has the same dynamic range as FP32 (8 exponent bits) with reduced mantissa precision. For inference on consumer and datacenter GPUs (which compute in BF16 natively), there is no meaningful quality difference.

Usage

Drop-in replacement for the original — just change the repo ID:

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

vae = AutoencoderKLWan.from_pretrained(
    "Calamdor/Wan2.2-T2V-A14B-BF16", subfolder="vae", torch_dtype=torch.float32
)
pipe = WanPipeline.from_pretrained(
    "Calamdor/Wan2.2-T2V-A14B-BF16", vae=vae, torch_dtype=torch.bfloat16
)
pipe.to("cuda")

output = pipe(
    prompt="Two anthropomorphic cats in comfy boxing gear fight intensely on a spotlighted stage.",
    negative_prompt="色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走",
    height=720,
    width=1280,
    num_frames=81,
    guidance_scale=4.0,
    guidance_scale_2=3.0,
    num_inference_steps=40,
).frames[0]
export_to_video(output, "output.mp4", fps=16)

Note: Requires diffusers installed from source until the next stable release includes Wan2.2 support:

pip install git+https://github.com/huggingface/diffusers

OneTrainer

This repo is compatible with OneTrainer for LoRA and fine-tune training of Wan2.2 T2V A14B.

License

Apache 2.0, same as the original. See Wan-AI/Wan2.2-T2V-A14B-Diffusers for the full license text and usage terms.

Credits

Original model by Wan-AI. See the original repo for the full technical description, paper, and citations.