Wan 2.2 LoRA API

Wan 2.2 LoRA animates an opening frame with Alibaba's Wan 2.2 image-to-video model and lets you load your own trained LoRA weights on top of it to steer style, character or motion. The frame and a motion prompt are required. LoRAs go into three optional slots with up to three entries each, and every entry is a direct URL to a weights file of up to 1 GB plus a strength from 0 to 4. Clips are exactly 5 or 8 seconds long, at 480p or 720p.

Three slots mapped to two denoising passes

Wan 2.2 denoises in two passes. high_noise_loras attaches weights to the high-noise expert, where composition and motion are decided; low_noise_loras attaches them to the low-noise expert, where detail and texture resolve; loras applies them to both. A line-art or grain treatment usually belongs in the low-noise slot, and a camera or movement style in the high-noise one.

Separate strengths for camera style and render style

Because the slots attach to different passes, a motion LoRA and a texture LoRA can share one request with their own scales, and you can turn one up and the other down instead of balancing a single blended weight. Inside a single slot, stacked LoRAs blend rather than layer, which is why adding them one at a time is the safer order.

Weights stay on your hosting

Each file is fetched from the URL you supply for the duration of the run, and nothing is stored on the platform. Direct download links work, as do Civitai model pages or download links and Hugging Face blob or resolve links, which are resolved to the file automatically. Other share pages and anything behind a login fail the fetch, so weights that must stay private belong behind a signed link.

The base model when the slots are empty

All three arrays default to empty. A request without LoRAs returns the base Wan 2.2 image-to-video result for the same frame and prompt, which is the reference that any strength sweep is measured against.

A house illustration style across a series

Train a LoRA on your line art and colour treatment and load it into the low-noise slot at a strength near 1. The opening frame supplies the character and set of each clip while the LoRA supplies the rendering, so every clip in the series shares it. The rooftop-walk example is the kind of starting point this needs: a cel-shaded character on a rooftop at night, run without a LoRA, with the prompt asking for the anime style to hold through a tracking move.

Camera style kept apart from render style

Put a motion or camera LoRA in the high-noise slot and a texture LoRA in the low-noise slot of the same request. When the movement is right but the texture is too heavy, lower only the low-noise scale and rerun with the same frame, prompt and seed.

Uncensored-style weights

Weights trained for uncensored styles run the same as any other file: nothing in the request declares content and no LoRA is inspected. What bounds the work is your rights to the weights, the training data and the opening frame, together with consent, age and local law for everyone depicted.

Finding the right strength

Hold the frame, the prompt and the seed, then sweep one LoRA's scale from 0.5 to 1.5 across four runs at 480p. Above roughly 1.5 the style tends to consume the subject instead of styling it. Keep the run where the style is present and the subject from the frame is still intact, and rerender that combination at 720p.

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "alibaba/wan-2.2-lora/image-to-video",
  "input": {
    "duration_seconds": 5,
    "image_url": "https://example.com/reference.png",
    "prompt": "A slow aerial shot drifting over a misty pine forest at sunrise",
    "resolution": "480p"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.