Character swaps often fail because they either destroy the background or drift in timing; this adapter explicitly trades model capacity toward keeping the source scene while injecting a reference character style. It’s a lightweight LoRA edit designed to be applied at inference to a MiniMax H3 Ref2VA runtime rather than a standalone replacement model.
Key Capabilities
- Targeted character replacement: conditions on a reference image to change a single person’s identity, outfit, and art style while attempting to preserve the original camera, lighting, background objects, and the movements of other people — so you can retarget a character without recreating the whole scene.
- Adapter workflow and compatibility: published as a single .safetensors LoRA (final checkpoint after 1,000 updates). Apply at strength 1.0 to a Ref2VA-capable MiniMax H3 base (VAEs and base weights must be provided separately) and run in ComfyUI or compatible runtimes — so it plugs into existing MiniMax H3 pipelines.
- Practical operating constraints: the training set is small (94 synthetic edit triplets, 40 unchanged clips; 76 edits used for optimization) and regularization clips are short, so short continuous shots (~4–5s at 24 fps) and conservative strength settings give the most reliable results.
- Known failure modes: temporal drift across longer windows, imperfect hard-cut handling, and inconsistent facial-expression fidelity; audio is not reliably preserved by the model and often requires postprocessing.
Who it's for and trade-offs
Great fit if you need quick, scene-preserving single-character swaps in short video clips and you already run MiniMax H3 Ref2VA in ComfyUI or a similar runtime. It’s useful for experimentation, style transfers, and localized edits where keeping the original background matters.
Look elsewhere if you need robust multi-character replacements, guaranteed lip-sync or audio fidelity, long uninterrupted sequences (>~10–15s) with strict camera-cut timing, or an adapter-free end-to-end model. The adapter inherits MiniMax H3’s licensing constraints (MiniMax H3 Community License) and requires the base weights and training assistant artifacts that are not bundled here.
