Realfencer IT2V Model (Leoma fine-tuned Wan 2.2 I2V)

Fine-tuned Wan 2.2 Image-to-Video pipeline for the Leoma Bittensor subnet (TI2V @ 832×480, 81 frames).

• Base: Leoma leader Wan I2V checkpoint
• Method: LoRA fine-tune (rank 64), merged into full weights
• Training checkpoint: step 500 (challenge profile, 2k clips)
• Deploy spec: 832×480, 81 frames @ 16 fps, guidance scale 5.0, 30 inference steps

Load with Diffusers

import torch
from diffusers import WanImageToVideoPipeline
from diffusers.utils import export_to_video
from PIL import Image

pipe = WanImageToVideoPipeline.from_pretrained(
    "Realfencer/IT2V_model",
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

image = Image.open("first_frame.png").convert("RGB").resize((832, 480))
prompt = "your motion prompt"

out = pipe(
    image=image,
    prompt=prompt,
    negative_prompt="low quality, blurry, distorted, flicker, temporal jump, identity drift",
    width=832,
    height=480,
    num_frames=81,
    guidance_scale=5.0,
    num_inference_steps=30,
    generator=torch.Generator(device="cuda").manual_seed(42),
)
export_to_video(out.frames[0], "out.mp4", fps=16)

Training metadata

See training_metadata.json for LoRA hyperparameters and deploy settings.

Bundled project code

The leoma_project/ folder contains the Leoma subnet tooling and fine-tuning scripts used to produce this model (no datasets or secrets).

VRAM

Full-precision inference at deploy resolution typically needs ~80 GB VRAM (production miner parity). Quantized/offload paths may work on smaller GPUs for local testing.