Summary
This tutorial from Veteran AI covers JoyAI Image Edit, a pair of diffusion models developed by JD that run inside ComfyUI — one for single-image editing and one for multi-image reference generation. Both models share the Wan 2.1 VAE but use separate text encoders, and the video focuses specifically on an optimized INT8 ConvRot version designed to fix a persistent pixel-alignment problem that plagues the official workflow.
The presenter walks through a series of increasingly complex editing scenarios: removing a person (including shadows) from a tiled background, swapping material textures on objects while preserving silhouette, repositioning items in a scene using annotated bounding boxes, and generating multi-angle views of a 3D object using yaw, pitch, and zoom parameters. A key section demonstrates why the default EmptySD3LatentImage resolution of 1024×1024 breaks pixel alignment for portrait images and presents two methods to fix it — scaling to preserve aspect ratio and a more robust tile-based approach.
For anyone working with ComfyUI image editing pipelines, the video is practical reference material: prompt structure guidelines (what to modify, what to lock, where to set boundaries), compatibility notes for extended nodes requiring the latest ComfyUI version, and live workflow demonstrations hosted on RunningHub make the techniques immediately reproducible.
📺 Source: Veteran AI · Published July 22, 2026
🏷️ Format: Tutorial Demo
