alibaba / wan 3.0 prime r2v

Overview

Wan 3.0 Prime R2V is Alibaba’s reference-to-video model in the Wan 3.0 series. You provide a detailed prompt plus optional references. The model blends them into a single MP4 video at 30 fps. It’s useful when you need stable identity, style, and voice across shots.

What you can build

• Character-consistent clips using a small set of reference images
• Product videos that match brand style from reference frames
• Motion-guided scenes using a short reference video as movement inspiration
• Dialogue-driven scenes that mimic a reference voice from an audio clip
• Multi-shot sequences written as a timed script inside one prompt

Inputs

• A text prompt that describes the scene, actions, and camera language. You can refer to references by index, like “Image 1” or “Video 1”.
• Optional reference images to lock subject identity or style. Provide up to 10 images in JPEG/JPG/PNG (no transparency), BMP, or WEBP. Each image must be 240–8000 px per side, up to 20 MB, and no wider than 8:1.
• Optional reference videos to guide motion, pacing, or framing. Provide up to 5 clips in MP4 or MOV. Total reference video time must stay within 15 seconds. Each clip must be at least 16 fps and no larger than 100 MB.
• Optional reference audio to guide voice timbre or soundtrack feel. Provide up to 5 WAV or MP3 files. Total reference audio time must stay within 15 seconds. Each file must be no larger than 15 MB.
• A target output duration in seconds. The generated video can be 2–30 seconds long. If you include reference video, keep input plus output within 30 seconds total.
• A resolution tier for the output video: 480p, 720p, or 1080p.
• An aspect ratio choice. “Adaptive” lets the model choose framing. Fixed ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16.
• An option to include or omit an audio track in the output video.
• An optional prompt expansion setting that rewrites short prompts to add detail.
• An optional random seed to help repeat a result range.

Outputs

• A generated MP4 video file.
• The output includes the final width and height, frame rate (30 fps), and the realized duration.
• The response also includes the seed used for the generation, which helps you rerun similar attempts.

Recommended settings

• For character consistency: use 3–8 reference images with varied angles and lighting.
• For motion control: add 1 short reference video and keep the output at 5–10 seconds.
• For short prompts: keep prompt expansion on, then edit the expanded prompt if needed.
• For iteration: start at 480p, then rerun at 1080p after you like composition.
• For repeatable experiments: set a fixed seed and change one input at a time.

Limitations

• Reference video and reference audio have a hard total limit of 15 seconds each.
• If you include reference video, input plus output duration must stay within 30 seconds.
• PNG transparency is not supported for reference images.
• A fixed seed can’t guarantee identical output across runs.
• Blurry, low-resolution, watermarked, or heavily compressed references reduce identity stability.
• Noisy audio, overlapping speech, or strong background music can degrade voice matching.
• Some providers return a download link that expires after about 24 hours. Save outputs promptly.
• Some user interfaces may force an “AI generated” watermark for compliance.

Safety & compliance

• Many deployments apply automated safety checks to both inputs and outputs.
• Don’t upload references you don’t have rights to use.
• Get consent before using a real person’s likeness or voice.
• Label AI-generated video where your policy or local law requires it.
• Avoid illegal content and prohibited harm, including impersonation for fraud.