Wan2.1-Fun-V1.1-1.3B-InP (Diffusers)
This is a diffusers-format conversion of alibaba-pai/Wan2.1-Fun-V1.1-1.3B-InP (Wan-Fun Inpaint V1.1 1.3B) from VideoX-Fun format.
Model Details
- Architecture: WanTransformer3DModel with
in_channels=36(16 noise + 4 mask + 16 image latent) - Parameters: 1.3B
- Pipeline:
WanImageToVideoPipeline(standard diffusers, no patching required) - Resolution: 480x832 (480p) recommended
- Frames: 49 frames at 16fps (~3 seconds)
This model has the same I2V architecture as the official Wan2.1-I2V-14B-480P (in_channels=36), but at 1.3B scale.
Usage
import torch
from diffusers import WanImageToVideoPipeline
from PIL import Image
pipe = WanImageToVideoPipeline.from_pretrained(
"engineerA314/Wan2.1-Fun-V1.1-1.3B-InP-Diffusers",
torch_dtype=torch.bfloat16,
)
pipe.enable_sequential_cpu_offload()
image = Image.open("first_frame.png").convert("RGB")
output = pipe(
image=image,
prompt="A person is talking naturally",
negative_prompt="static, blurred, low quality",
height=480,
width=832,
num_frames=49,
num_inference_steps=50,
guidance_scale=5.0,
)
from diffusers.utils import export_to_video
export_to_video(output.frames[0], "output.mp4", fps=16)
Conversion Details
Converted from VideoX-Fun format using 1:1 weight key mapping (983 keys). No architectural modifications were needed -- the standard WanImageToVideoPipeline handles in_channels=36 natively.
Components
| Component | Source |
|---|---|
| Transformer | Converted from alibaba-pai/Wan2.1-Fun-V1.1-1.3B-InP |
| VAE | Wan-AI/Wan2.1-T2V-1.3B-Diffusers |
| Text Encoder | Wan-AI/Wan2.1-T2V-1.3B-Diffusers (UMT5-XXL) |
| Image Encoder | Wan-AI/Wan2.1-I2V-14B-480P-Diffusers (CLIP ViT-H-14) |
| Scheduler | UniPCMultistepScheduler (flow_shift=3.0) |
Comparison with TI2V variant
| This model (InP) | TI2V | |
|---|---|---|
in_channels | 36 (noise + mask + image) | 32 (noise + image) |
| Pipeline patches | None needed | prepare_latents override required |
| Origin | Wan-Fun Inpaint | Wan-Fun Camera Control (adapter removed) |
Acknowledgements
- Alibaba PAI / VideoX-Fun for the original Wan-Fun models
- Wan-Video for the Wan 2.1 architecture