Alibaba: Wan 2.6

Alibaba Wan 2.6 is a video-generation family with distinct text-to-video, first-frame image-to-video and reference-to-video profiles. The maker documents synchronized audio and multi-shot narratives. A written scene can become a clip, a prepared still can establish the opening or visual references can guide characters and objects in a new scene. The useful output is the completed short video and its sound, prepared for review or use in an edited sequence.

For text generation, describe the subject, environment, movement, camera and sound. The legacy 2.6 text guide accepts Chinese or English prompts up to 1,500 characters, with a separate 500-character negative prompt; longer text is truncated. For a product animation, follow a toy delivery van crossing a workbench, pause beside a small parcel and end on the parcel as a quiet mechanical sound marks the stop. Keep the sequence simple enough that its central action remains easy to judge.

Multi-shot control is a specific part of the 2.6 workflow. The maker’s text guide offers single or multi shot_type settings, with single as the default. In that text mode the setting takes effect with prompt rewriting enabled, and the parameter has priority over contradictory wording in the brief. Describe the shots and timing consistently with the selected setting. The later Wan 2.7 guide instead interprets shot structure directly from the prompt, so that newer preparation should not silently replace this edition’s documented controls.

First-frame generation uses an image to establish the opening and a prompt to describe what follows. Prepare the intended crop, subject and visual details in the still before asking for motion. The maker’s comparison identifies first/last-frame and source-clip continuation as additions in the 2.7 image workflow, rather than controls to assign to 2.6’s first-frame profile. For a product animation, review the object through the whole movement and inspect whether its shape, markings and scene remain suitable for the intended shot.

The current legacy 2.6 reference API accepts images and videos through an ordered reference list. It permits up to five images and three videos, with a combined maximum of five files. Each reference should contain one subject. The first maps to character1, the second to character2 and so on; use those identifiers in the prompt rather than borrowing Wan 2.7’s type-specific labels. Describe each character’s action and lines clearly so a multi-character interaction can be checked for attribution and identity.

The reference guide lists JPEG/JPG, PNG without transparency, BMP and WEBP images up to 20 MB, with each dimension between 240 and 8,000 pixels. Reference video may be MP4 or MOV up to 100 MB; the current reference_urls field accepts one-to-thirty-second videos. The deprecated video-only field has a different minimum duration and is not the same preparation contract. Choose the source passage and subject that should inform the scene, and keep the reference roles explicit. A recognisable character can still be placed or voiced incorrectly in the result.

The maker’s text generation supports supplied audio through its documented audio source field, and the output modes include sound synchronized with the picture. State the intended dialogue, performance and atmosphere in the brief and review the actual words and effects. The reference API’s optional audio toggle is specifically labelled for the Flash reference variant; it is not a universal silent-video control to assume for every 2.6 mode. Sound and picture both need inspection against the intended scene.

Current maker specifications list 720p and 1080p output, with two-to-fifteen-second text and image generation and two-to-ten-second reference generation. Its video guides describe MP4 at 30 fps, while the reference API selects an explicit width-by-height size rather than a bare ratio label. Landscape, portrait and square options have their own documented dimensions. Keep the selected mode’s duration and size rules with the brief instead of borrowing Wan 3.0’s thirty-second generation allowance or a different edition’s adaptive configuration.

Generation is processed as a task, and a successful result supplies the video location. The maker documents twenty-four-hour result-link validity. Retain the actual clip, source references and prompt for the editing review. A fixed seed can help compare requests, but the legacy guide explicitly says it does not guarantee identical results. Review continuity, reference identity, object contact, requested shot structure, dialogue and visible details. The next correction should address the specific problem found while preserving the successful direction of the shot.