This workflow performs reference-based video-to-video editing with LTX 2.3, steered by the Union Control IC LoRA and canny edge guidance. You supply a single edited frame from your clip (for example, painting a red shirt onto a boxer), and the pipeline propagates that visual change to every frame in the video while preserving motion and structure. The core processing happens in the custom processing block (UUID: 9278a6c3-b46a-4e21-8d09-5103f85ffec6), which runs LTX 2.3 inference with the Union Control IC LoRA and integrates canny control to keep edges and silhouettes consistent across time.

Technically, LoadVideo ingests your source clip and feeds frames into the inference block, where canny edge maps are used as a structural prior so the subject’s pose and scene layout remain stable. LoadImage brings in your manually edited reference frame, which the IC LoRA uses as an appearance guide to replicate the change throughout the sequence. After processing, VHS_VideoCombine assembles the generated frames back into a video, and SaveVideo writes the final output to disk. This approach is ideal when you want targeted, consistent replacements—like color swaps, wardrobe fixes, or logo/graphic updates—without re-shoots or heavy roto work.