Summary
This tutorial from the Veteran AI channel introduces JoyAI-Echo, an open-source long-video generation model from JD.com’s research team built on the LTX 2.3 foundation. Unlike standard text-to-video models limited to roughly 10-second clips, JoyAI-Echo stitches together up to 15 individual shots — each 241 frames at 25fps, approximately 9.6 seconds — producing videos up to 2 minutes and 24 seconds with synchronized audio generation. A cross-shot memory mechanism maintains character, environment, color, and audio consistency across segments, addressing the central challenge of long-form AI video: coherence over time.
The tutorial covers two ComfyUI deployment paths in technical depth. The lossless version preserves full model precision using an official extension but demands serious hardware: the main model weighs 46GB and the text encoder over 20GB, requiring an A800 80GB GPU. The creator demonstrates a pre-configured cloud instance on Youyun Zhisan (Compshare) to make this path accessible. The simplified GGUF quantized version supports consumer GPUs with far lower VRAM requirements but sacrifices audio quality and generation stability.
Two lossless workflows are demonstrated: a batch mode for pre-planned multi-shot sequences using the JoyEcho_Generate node, and a shot-by-shot workflow that passes memory state forward between segments via JoyEcho_SingleShotGenerate — enabling incremental review and adjustment. The creator recommends keeping shot counts under eight for maximum consistency, and notes that the Gemma-3-12B text encoder requires absolute file paths. Example outputs include a Rhine River travel vlog and a Princess Elsa narrative video with multiple characters and scene transitions.
📺 Source: Veteran AI · Published June 23, 2026
🏷️ Format: Tutorial Demo
