🎤 LongCat-Video-Avatar 1.5: Audio-Image-to-Video

A reference implementation of checkpointed inference for long-running AI workloads

Many AI models — video generation, audio generation, and other long-running inference workloads — need more processing time than a single available GPU session provides. This Space, inference-checkpoint-saver, is a reference implementation of a general-purpose technique for that situation: checkpointed save & resume inference, where a model's generation state is written out mid-process and can be reloaded later to resume from the exact point it left off, with no loss of progress and no discontinuity in the output.

We use LongCat-Video-Avatar-1.5-2nd, an audio-driven video generation model, as one concrete example implementation of this approach. While processing is underway, the full generation state — accumulated video latents, the decoder's cache, the current diffusion step, and the reference inputs — is serialized to a portable state file. That file can be loaded in any later session, and generation continues exactly where it left off, frame by frame, with no skipped or discontinuous output. This is what makes it possible to produce up to 2 minutes of audio-driven video from a single reference image, well beyond what one continuous processing session would otherwise allow.

This pattern generalizes to any autoregressive or chunked generation model that needs to produce long-form output under a limited amount of continuous compute time, not just this one.

Upload a reference image, a driving audio clip, and a short text prompt to get started.