JoyAI-Echo is an open-source long-horizon audio-visual generation project from JD.com, titled "Long-Horizon Audio-Visual Generation for Persistent Stories and Interactive Worlds," targeting long-horizon audio-visual generation in persistent stories and interactive worlds. The repository maintains two mutually independent projects: Echo-WM and JoyAI-Echo 1.5. The two have their own environments, checkpoints, and entry points, and do not share a Python environment or checkpoint directory. echo_wm/ comes with its own copies of ltx-core and ltx-pipelines, so installing one project does not affect the other. Checkpoints must also be downloaded separately.
In terms of use, the project mainly targets scenarios requiring long time spans and joint audio-visual generation, such as persistent storytelling and interactive world simulation.
In terms of main capabilities, according to the repository description, JoyAI-Echo 1.5 / Echo-LongVideo, released on 2026-08-28, includes long-horizon generation, consumer-GPU inference profiles, and Director Agent. Previously, on 2026-08-26, Echo-WM was released, an omnimodal world model for interactive audio-visual generation. On 2026-06-22, JoyAI-Echo 1.0 was released and is now kept in the echo1.0 archive branch.
Echo-WM's roadmap shows: it is currently based on LTX-2.3, including a bidirectional audio-visual DiT (Base, about 10 seconds) and the current public preview Flash Preview / Causal (chunked causal attention, KV-cache rollout, 4-step inference); future plans are to migrate Base and Causal to LTX-2.5, and to reduce the cost of long rollouts through sparse attention, more compact caching, and the runtime stack. Other plans include sparse/low-bit-width kernels such as SageAttention, fused attention such as FlashAttention/FlashInfer, paged KV-cache, FP8/TensorRT, etc., all of which are plans or directions and do not represent something already implemented.
In terms of usage, the environment must be configured, checkpoints downloaded, and the corresponding entry point run separately according to each project's READMEthe two projects have isolated environments.
The applicable scenarios are academic research and non-commercial use. The project is based on LTX-2 from Lightricks Ltd., and JD.com has modified some of its code, for academic and research purposes only; commercial use of LTX-2 or its derivatives requires contacting Lightricks Ltd. The project retains all of LTX-2's copyright, license, patent, trademark, and attribution notices, and continues to be bound by the LTX-2 community license agreement. If used for research or products, the authors provide citation instructions.
In terms of limitations, it is explicitly limited to academic research and non-commercial use; the two subprojects have independent environments and checkpoints and do not share them; some roadmap items (LTX-2.5, sparse attention, paged KV-cache, FP8/TensorRT, etc.) are future plans. Specific installation commands, model file lists, hardware thresholds, and performance data are not yet provided in the official text and require consulting each README.
