Wan 2.6

Text-to-video and image-to-video support video durations up to 15 seconds with custom aspect ratios, while reference-to-video uses up to three video clips for subject consistency and is capped at 10 seconds. Text-to-video creates footage purely from written prompts, image-to-video anchors the initial aesthetic to a single keyframe, and reference-to-video preserves recurring people, animals, or objects across different scenes.

Use the reference-to-video variant when recurring characters or specific objects need to appear across multiple shots. By providing up to three source videos (at 16 FPS or higher) and tagging them as @Video1, @Video2, or @Video3 in your prompt, the model locks facial geometry, wardrobe, and identity while placing subjects in new environments and camera setups.

Multi-shot scheduling directs multiple camera angles and transitions in a single generation using bracketed timestamps like Shot 1 [0-5s] and Shot 2 [5-10s] directly in the prompt. With multi-shot mode enabled, the model plans coherent shot transitions, preserving lighting, character appearance, and environmental continuity across runs up to 15 seconds.

Wan 2.6 synthesizes speech, ambient sound effects, and frame-accurate phoneme lip synchronization directly during video generation. Because audio and video are generated natively together at 24fps, dialogue matches mouth movement without the drift or uncanny artifacts typical of secondary dubbing tools.

Wan 2.6 is engineered for physical realism up to 15 seconds, so highly abstract or non-physical surreal scenes may be coerced into grounded real-world physics. To get the best results, avoid radical environment shifts between shot brackets, provide clean audio stems when supplying custom audio, and disable prompt expansion when executing strict multi-shot timing syntax.