Reference-Guided Video Continuation feature

Dear SeaArt Team

It is good to see the new scene-guidance possibilities offered by the “Reference to Video” function, for instance in your Wan 3.0 and MiniMax H3 related models.

However, generating two or three connected clips that form one continuous scene still faces a major limitation: prolongation based mainly on the final frame of the preceding clip tends to induce substantial drift in reasonably complex / filmic scenes, while genuine persistent scene memory across clips would be the natural way forward – but is not yet available.

A practical intermediate solution might be to allow users to provide a separate set of Previous-Clip References, in addition to the reference images for the clip currently being generated. The former would serve as context anchors only: user-selected frames showing what has already been established in the preceding clip – characters, clothing, objects, spatial configuration, camera perspectives, etc. – without being treated as content or keyframes that should appear in the new clip.

The user would deliberately select these frames and describe, where useful, their chronological or spatial relation. A camera orbit is a simple example. Suppose Clip 1 moves roughly halfway around a scene. The user could provide several representative frames along that trajectory and indicate their approximate camera positions. One suitable frame near the end of Clip 1 – not necessarily its literal final frame – could then serve as the start frame for Clip 2. The new clip would continue the orbit from there, while the other Previous-Clip References provide visual information about parts of the room, characters, clothing, or objects that may no longer be visible in that start frame. Crucially, these frames describe the preceding clip; they are not frames that the new clip should reproduce.

In other words, the user manually provides and structures part of the scene memory that a fully persistent system would otherwise have to infer and retrieve automatically.

My concrete feature request would therefore be to add a second, clearly separated reference input for video continuation: Previous-Clip References – user-selected frames from the preceding clip that condition the continuation but are explicitly excluded from the timeline to be generated.

From a UI perspective, it might therefore make sense to separate the Previous-Clip Context (verbal description + reference images) from the actual generation prompt (generation instructions + new reference images), with each field having its own associated reference images and text description. Besides keeping “what has already been established” separate from “what should happen next”, this would also keep image referencing simple and unambiguous: @Image1 in the Previous-Clip Context field would refer to the first reference from the preceding clip, while @Image1 in the actual generation prompt would refer to the first reference image for the new clip. This avoids having to mix two sets of images serving two different purposes within a single prompt. The exact UI implementation is of course secondary; the key requirement is a clear separation between previous-scene context and current-clip generation.

This new Reference-Guided Video Continuation feature could provide a relatively lightweight bridge between current last-frame prolongation and future persistent scene-memory systems, while leaving the difficult-to-automize selection, retrieval, and description of the relevant previous-scene information to the user.

As a related feature request, allow users to specify whether each uploaded image for the clip to be generated is a start frame, end frame, or reference frame. (While there can be multiple reference frames, the start frame and end frame must each be unique.) At present, users have to choose between the “Reference to Video” option and the Start/End Frame option. As the example above illustrates, it is essential for this approach – but also useful for one-off clips – that users can explicitly assign these different roles to individual images within the same generation.

P.S. For simple linear clip prolongations, the current prolongation functionality already in place is fine. For filmic / scene-heavy scenarios that tell a coherent story, which needs to be consistently visualized, this is where this feature would probably shine.

Thanks for sharing this request with the development team, if you think it could help to improve SeaArt’s consistent-multi-clip generation capabilities.

Best regards,

Scotti