Alibaba: Wan 2.7

Alibaba Wan 2.7 is a video-generation family with named text-to-video, image-to-video and reference-guided profiles. The manufacturer documents audio-synchronized scenes and multi-shot narratives, while its image workflow adds first/last-frame generation and continuation of an initial video. The useful result is a short clip with its sound, such as a product transition, a compact story or a continuation that fits an existing shot. Select the mode from the source material and the control needed over the beginning or ending.

For text-to-video, describe the subject, setting, action, camera, style and sound. A storyboard can identify each shot and time range, while a single-shot brief should say that the action remains continuous. The Wan 2.7 prompt guide lets the model interpret shot structure from the written brief rather than using 2.6’s shot_type control. The current text API accepts Chinese or English prompts up to 5,000 characters, with a separate 500-character negative prompt; excess text is truncated. Keep essential action and constraints clear within that allowance.

First-frame image-to-video animates a supplied opening image. First/last-frame generation also defines the endpoint and the motion connecting it. The 2.7 image guide supports a driving audio source alongside those frame combinations. Prepare endpoints that share a scene and an intended subject, and describe the transition rather than asking the model to reinterpret both images as unrelated style references. The output approximates the first image’s proportions; the guide treats resolution as a pixel-count tier and rounds encoded dimensions appropriately.

Video continuation begins with a source clip and can also use a closing image. The requested duration is the total returned video, including the original segment. The guide’s example combines a two-second source with ten seconds of new content into a twelve-second result. Prepare the source’s final action and say what happens next, so the camera and subject continue in a coherent direction. Review the join for appearance, movement, light and sound rather than assessing only the newly generated ending.

The maker’s current separate reference profile uses images or video to guide characters, objects or a scene. Its guide allows up to five reference images and videos in total and uses type-specific labels such as Image 1 and Video 1. Describe the role of each material and the intended interaction. For a reference-guided scene, show a character from one reference carrying the object from another through a specified location. Reference framing differs from first-frame generation: the brief can establish a new composition instead of requiring the video to open on one supplied still.

For custom audio in text generation, the documented formats are WAV and MP3, with a two-to-thirty-second duration and a 15 MB maximum. The selected video duration determines the portion used: longer audio is clipped to that duration, while audio shorter than the video leaves the remainder silent. Prepare the exact audio passage that belongs to the shot and leave room for the intended visual action. Without a supplied audio file, the maker describes generating background music or effects suited to the scene; inspect the actual sound and any requested speech with the picture.

The current model matrix lists 720p or 1080p MP4 at 30 fps, with two-to-fifteen-second text and image output. The separate reference profile is limited to two to ten seconds. Text generation offers 16:9, 9:16, 1:1, 4:3 and 3:4 framing. Image-conditioned tasks use their own approximate source-proportion rules, so prepare the input crop for the intended destination. A duration or aspect rule from the reference profile should not be substituted for the frame or continuation workflow.

The documented maker flow is asynchronous: a task identifier tracks processing, and a successful result supplies the video location. Results are documented with twenty-four-hour link validity, so retain the actual clip and the creative materials for editing review. Prompt rewriting, watermark and seed options belong to the selected interface. The maker says a fixed seed can improve reproducibility but does not guarantee an identical result. Compare the returned artifact with the brief rather than assuming settings alone preserve every detail.

Compared with Wan 2.6, the 2.7 image guide specifically adds first/last-frame control and video continuation through multimodal media inputs. Wan 3.0 introduces a different all-in-one structure and longer generation allowance. This 2.7 profile keeps its task-specific controls and limits. Review the whole scene for subject continuity, frame anchors, reference identity, dialogue, motion and joins, and make the next correction specific to the actual issue found.