H3 character swap LoRA vs the H3 Max Recast API: which to use

The open H3 Character Swap LoRA from Akatz Labs is a 148 MB adapter you run yourself, and its own notes say 4-5 second clips are the sweet spot and longer ones drift, with a 15-second test degrading to a blurred frame. H3 Max Recast, the hosted route that Sume lists as h3-max-recast, takes 5-30 second sources with up to 15 seconds per shot and keeps the original cuts and audio. Pick the LoRA to experiment on your own GPU; pick Recast when the deliverable is a 15-30 second ad with cuts.

Both target the same job, replacing a person in a video, so buyers now meet them side by side.

What each one is

The LoRA is a rank-16 adapter for MiniMax H3 Ref2VA released on September 25, 2026, with open weights on Hugging Face, plus the training dataset and configuration. It takes a reference video, a character image or sheet and a prompt, and swaps one person while keeping the scene, camera and props (ComfyUI Wiki).

H3 Max Recast is a hosted model: one source video, one photo per new person, an optional prompt, 768p or 1080p, and the output keeps the source motion, camera, cuts and audio (fal H3 Max Recast).

When the open adapter is the right call

If you are prototyping character looks, testing stylised or illustrated characters in live action (which the LoRA page highlights), or need weights you can inspect, run the LoRA. You pay in setup and GPU hours and you own the failure modes: the page itself warns that close-up expressions often do not follow the original performance.

When a hosted swap is simpler

A finished ad has cuts, sound and a deadline. Recast handles up to 30 seconds with cuts, so the work is cutting the source at shot boundaries and sending photos. On Sume the request is one source video_url, 1-4 reference_image_urls, the inspected duration, and an optional prompt; the price is $0.375 a second at 768p and $0.5625 at 1080p, and the result is mirrored to media.sume.com.

Sume's catalog is a fixed list of model ids you can read from GET /v1/video-router/models, so a community LoRA is not something you can select there. If your workflow depends on a specific adapter, run it yourself and use Sume for the stages around it.

• Keep a 5 second test slice: it is the cheapest valid Recast job and shows how the face holds before you pay for 30 seconds.
• Whichever route you use, get the written consent of anyone whose likeness you swap in.