Changing a character across an ~hour long TV episode

I’ve been working on a repeatable process to do character swaps (& other types of changes) across a complete TV episode / movie. It’s working pretty well now.

I thought this was just funny… not so much a true “swap” but asking the character to dance in every scene. Choreography courtesy of an LLM. If you’ve seen the show, it’s Serious Detective Drama (TM) so the fact the main character suddenly starts pumping his arms or sliding into every scene with a dance move made me laugh out loud when I saw some of the generations.

Details on the workflow:

Model: MiniMax H3, first-and-last-frame mode (FL2VA), in ComfyUI. 8-step turbo LoRA, generated at 1024×512 for the 2:1 picture.

First and Last: Every generated pass starts and ends on an *original* frame from the episode, so the clip lands back in the real footage with no visible join. Longer scenes are chained passes (about 2–7 s each)

Audio: the episode’s own audio into H3’s audio guide, so the generation stays consistent with what’s being said; the dialogue and soundtrack you hear are the untouched originals.

Prompt: Who he is and what he’s wearing, said once, then the choreography for the pass – courtesy of Claude.

Multi-scene: 42 scenes – started with the full episode, then had a model run through it to identify scenes with the main character that would work for the swap (not too distant, few fast cuts, etc.) & identify start and finish, with a ChatGPT harness.

Post processing: Basically none, I did go through and rate all the clips in a rating process I’ve setup, most of these were 5* and felt consistent.

I’ll share a bit more on the workflow / process if people are interested.