alibaba / Wan 2.7 Reference

## Overview Wan 2.7 R2V by Alibaba generates short videos from visual references plus a written prompt. It uses your reference images or clips as identity anchors, then creates new motion and camera moves that match your description. You can also edit an existing clip by describing changes in plain words, with an optional reference image for guided edits. This helps you keep the same character or product look across scenes without manual keyframing. ## What you can build - Character-consistent clips from 1 to 3 reference images - Product shots that keep packaging and color details stable across scenes - Brand mascot animation that stays recognizable across multiple generations - Instruction-based edits of a short clip, like style changes or scene rewrites - Element replacement using an extra reference image, like swapping clothes or props - Social-friendly vertical videos by selecting a 9:16 aspect ratio ## Inputs - A required text description of the scene, action, and camera behavior, up to 5,000 characters - An optional negative description of what must not appear, up to 500 characters - A source video to edit when you want video editing mode (MP4 or MOV, 2 to 10 seconds, up to 100 MB) - Reference images when you want reference-to-video mode (JPEG/JPG/PNG/WEBP). You can provide up to 3 images. - A target output duration between 2 and 10 seconds - A resolution tier selection: 720p or 1080p - An aspect ratio choice (16:9, 9:16, 1:1, 4:3, 3:4), or auto selection based on your first reference image. When you edit a video, the aspect ratio follows the input clip. - An audio choice when a source video is provided: keep the original audio or generate a new audio track - An optional prompt expansion toggle that rewrites short prompts into longer scene directions - An optional watermark toggle that adds an "AI Generated" mark in the lower-right corner - An optional seed number for repeatable results (0 to 2,147,483,647) ## Outputs - A generated MP4 video file - Video metadata that describes the clip, such as width, height, fps, duration, and frame count - The seed used for the run, so you can recreate or iterate on the same motion - When prompt expansion is enabled, the final rewritten prompt may be returned alongside the video ## Recommended settings - For the most stable identity match, use 1 clear reference image and set duration to 2 to 5 seconds - For maximum detail, pick 1080p when your scene has faces, hands, or product close-ups - Use multiple references only when they show the same subject with consistent styling ## Limitations - Identity can drift in busy scenes, especially with multiple moving subjects - Fine details like small text, logos, and tiny accessories may change across frames - Long or highly choreographed actions can reduce resemblance to the references - Conflicting references can confuse the model and cause mixed features - Low-quality inputs reduce results. Blurry images, heavy compression, fast cuts, and noisy frames can cause flicker and artifacts. - Editing mode only accepts short clips within the supported duration and file size limits ## Safety & compliance - This model can apply content moderation to both inputs and outputs, which may block some requests - Only upload videos and images you have rights to use - Don’t generate or edit content that impersonates real people without consent, especially for deceptive or harmful uses - Follow Wiro’s content rules for prohibited or restricted content