A LoRA fine-tune of Wan-AI/Wan2.1-I2V-14B-480P-Diffusers on the REA aesthetic — mysterious, otherworldly, far-out, uncivilized; earth-resembling but slightly alien; atmospheric environments with weathered architectural form and a sense of isolation and vastness.

Generates 49-frame, 832×480, 16fps clips from a first-frame image + caption. Trigger word: REA (prepend to every caption at inference time).

Quick usage

python

import torch
from diffusers import WanImageToVideoPipeline
from diffusers.utils import export_to_video, load_image

pipe = WanImageToVideoPipeline.from_pretrained(
    "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers",
    torch_dtype=torch.bfloat16,
).to("cuda")

pipe.load_lora_weights("whp-rea/rea-wan-i2v-lora-v1", adapter_name="rea")
pipe.set_adapters(["rea"], [0.85])

image = load_image("your_first_frame.png")  # 832×480

video = pipe(
    prompt="REA A vast desert badlands at dawn, soft cool sky over rolling eroded sand-and-clay ridges...",
    image=image,
    num_frames=49,
    height=480, width=832,
    num_inference_steps=30,
).frames[0]
export_to_video(video, "rea_output.mp4", fps=16)

LoRA scale around 0.75–0.95 works well. Lower for subtler style transfer, higher to lean fully into the aesthetic.

Files

pytorch_lora_weights.safetensors — the final LoRA at step 300, ready to load via WanImageToVideoPipeline.load_lora_weights() (~210 MB)
lora_weights/000300/pytorch_lora_weights.safetensors — same file (finetrainers output convention)
lora_weights/000150/pytorch_lora_weights.safetensors — intermediate checkpoint at step 150, for comparing transfer strength

Training data

whp-rea/rea-world-clips-v1 — 25 deterministic camera traversals through 5 Gaussian-splat worlds generated in World Labs Marble, each captioned in REA literary style. Five trajectory primitives per world: walk_short, walk_turn_walk, look_around, pitch_sweep, figure_8. Bundles include first frame, video, pose/action arrays, and intrinsics (action conditioning unused for this Phase 1 LoRA — see "Phase 1 scope").

Training recipe


Base model	`Wan-AI/Wan2.1-I2V-14B-480P-Diffusers`
Method	LoRA, rank 16, alpha 16
Target modules	`blocks.*(to_q\|to_k\|to_v\|to_out.0)` (attention projections only)
Steps	300
LR	1e-4, constant_with_warmup, 30 warmup steps
Optimizer	AdamW (β1=0.9, β2=0.99, wd=1e-4, ε=1e-8)
Batch size	1 (no grad accumulation)
Resolution	49 frames × 480 × 832
Precision	bf16 + grad checkpointing + VAE slicing/tiling
Hardware	1× NVIDIA A100-SXM4-80GB
Trainer	huggingface/finetrainers (with bypass patches for torchcodec/decord, `--validation_dataset_file` deliberately omitted to avoid imageio dependency in inference path)
Trigger token	`REA` (prepended to all training captions via `id_token`)
Final training loss	0.036 (avg of last few steps, batch size 1 noise typical)
Wall-clock training time	1h 45m

Phase 1 scope

This is a text+image → video LoRA. The camera poses, action arrays, and intrinsics that live in the source dataset are not consumed by this fine-tune; they sit colocated in the bundles for forward compatibility with a future pose-conditioned model (LingBot-style). Phase 1 proves that the REA aesthetic transfers cleanly through Wan-family I2V; Phase 2 will explore action conditioning.

Limitations

Trained on 25 clips → distribution is narrow. Best on prompts close to the five training world types (desert badlands, sandstone tunnel, weathered stone gateway, ruined colonnaded terrace, redwood forest road).
Wan2.1-I2V-14B-480P is fixed at 832×480 / 49 frames natively; departures from that bucket may degrade.
No pose / camera control — camera motion is whatever the base model decides given the prompt + first frame.

Provenance

Trained via ML Intern on Hugging Face Jobs (job 6a19fe2a3a4b8cae6044e206) on 2026-05-29. Dataset, training script, and recipe details in the parent REA Splat project.