Player not loading? Watch on YouTube
This tutorial explains a local AI VFX pipeline in ComfyUI built around a merge of SkyReels V3 R2V and Wan 2.1 VACE. The speaker uses masks to define replacement areas, reference images to guide appearance, and ControlNets to constrain scene structure. Examples include environment replacement, character transformation and object removal.
The preparation workflow extracts depth maps, Canny edges and poses, tracks the camera, and isolates a subject from a text prompt. Viewers learn how to combine those outputs into a driving video and choose which parts the model can change. Blender geometry can supply depth sequences for a designed scene. Qwen Image Edit provides a local reference-image workflow; the speaker also uses the paid, closed-source NanoBanana.
Setup covers missing custom nodes, model downloads and input settings. The speaker recommends FP8 when hardware permits, with smaller GGUF versions for GPUs with less VRAM at some cost to quality. The basic workflow supports up to 121 frames, roughly five seconds, at a maximum of 720p. Advanced Patreon workflows add segmented generation and prompt scheduling.
The shot breakdowns show practical limits: a full-frame mask improves interaction effects but changes the subject's appearance, while references and mask shapes can push motion or effects in unwanted directions. Speed advice includes SageAttention, with RunPod as an optional rented-GPU route.
