Skip to content

← The Lab

OPENjutsu: Comfy open-sources a character-swap app on MiniMax H3, and a fully local fork follows within hours

Comfy published an MIT-licensed, open take on the 'swap the person in a clip' trick that Higgsfield's Genjutsu made popular. It runs on MiniMax H3 and SAM 3.1. A community fork already runs it entirely on a 16 GB card.

On Oct 26 at 4:14 PM MDT, @ComfyUI announced OPENjutsu, an open-source app that replaces a character in an existing clip while keeping the original motion, camera and timing. The post passed 450 likes, 300 bookmarks and 26K views overnight. The repo (Comfy-Org/comfy-openjutsu, MIT) wraps the akatz-labs H3 character-swap workflow, the same approach we covered from MDMZ earlier this month. It uses SAM 3.1 to mask the subject and MiniMax H3 to re-render it.

How the official version runs. According to the README, OPENjutsu deploys to the Comfy Developer Platform and runs on a 96 GB RTX PRO 6000. A warm 5-second, 480p part takes about two minutes, and you pay in deployment credits. Clips are processed in 5–15 second parts at 480 or 768. There's an optional step that makes the first frame of the new character with GPT Image 2.5 through Comfy Router. You can also try it as an app on Comfy Hub without setting anything up.

The local fork. Within hours, @Devozikjr published D3voz/openjutsu (also MIT), which swaps the hosted pieces for local ones: MiniMax H3, SAM 3.1 and Qwen3-VL 8B for captioning. Per its README, it offers a simple mode, an advanced mode for one to four characters, object swaps, swapping from a reference picture with SolarRenderer, and passes to fix faces and colours. The author lists an RTX 4060 Ti 16 GB with 64 GB of system RAM as the tested setup. Both projects arrive during the #OpenSourceHiggsfield challenge, which closes Oct 30 at 10 AM MDT and has an open-weights-only track.

Use it responsibly. Swap characters only in footage you shot or have permission to use, and only put in people who have agreed to it, or fictional and original characters. Label the result as synthetic when you publish. MiniMax H3 Community License terms also apply to the weights.

Our take: this is now a standard, free building block instead of a paid platform feature. For agencies, the useful version is the boring one: re-casting your own previs or stand-in footage with a licensed or original character before a paid shoot. Start with the Hub app to judge quality, then move to the local fork if you have a 16 GB card and want unlimited drafts.

Related tools on ReelStack

Skills & repos in our directory

More from the Lab

Topics: ComfyUI · Higgsfield · Open-weight video & audio models