Hi Anna team,
Thank you for opening video generation to Apps and Agents! We are integrating it into Timeline Studio / ChatCut, and image references are a useful first step.
Does Anna currently expose reference-video inputs through another API or SDK path? In the installed @anna-ai/app-runtime@0.20.0 contract we found imageUrl, but have not found a video-reference field. If this is not available yet, could you consider combined image + reference video inputs for models that support them?
The workflow we want to build
- Upload or generate a character image.
- Optionally provide a background image and confirm the starting composition.
- Upload a short motion reference, such as a dance or character interaction.
- Optionally derive a depth video locally.
- Generate a new clip using the chosen character/scene while following the reference motion, timing and framing.
A single still cannot communicate an entire motion sequence. Original video can guide actions and expressions; depth video can provide spatial/contour guidance. We do not assume that all video models support depth control, or that prompting alone guarantees exact reproduction.
Concrete example
The attached silent GIF preview of a 14-second side-by-side video shows the color motion reference on the left and a generated character replacement on the right. It illustrates the kind of temporal guidance we want to expose in our editor. This example was produced through a separate video-reference-capable generation path, not Anna’s current video API, and is not a claim of frame-perfect matching or perfect expressions.
Provider capabilities worth exposing
These are provider/model capabilities, not claims about what Anna currently exposes:
- Seedance 2.0: the official Volcano Engine “Doubao Seedance 2.0 Prompt Guide” explicitly describes combined image/video/audio references, including motion transfer and voice reference.
- MiniMax H3: the official MiniMax announcement “MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities” demonstrates video camera-motion reference + character image + matching audio in one context. This should not be conflated with H3 Max.
- H3 Max on fal: the current image-to-video schema exposes first/last-frame images and
target_audio_url, described as pinning/replacing the output soundtrack with the supplied audio. That is different from general audio-conditioned motion/voice reference; this endpoint does not list a video-reference input. - Kling VIDEO 3.0 Omni: the official “Kling VIDEO 3.0 Omni Model User Guide” documents video reference and voice-bound character elements. This is not evidence of an unrestricted generic music/audio-reference field in every Kling API. The guide also lists restrictions when combining video input and native audio; please expose actual endpoint/version constraints.
What would help
- Allow character/image references together with a motion-reference video and optional audio, where the selected model/endpoint supports that combination.
- Distinguish audio roles: voice/timbre reference, music/rhythm guidance, original-video audio preservation, and exact soundtrack replacement. Native audio generation alone is not audio-reference support.
- Expose model-specific supported reference types: original video, depth, pose, etc., and their intended control roles where supported.
- Document formats, duration/file-size limits, reference counts, and whether reference audio is used.
- Include these capabilities in model discovery, with explicit errors for unsupported inputs and CU estimates where available.
- Support uploaded storage files or signed URLs so apps can submit user-selected clips without requiring public hosting.
If there is already a supported route, a minimal SDK example would be very helpful. Otherwise, this would unlock motion-guided recreation workflows beyond image-to-video. Thank you!
