Most model merges either fine-tune or average weights; this release instead treats H3 as an architectural substrate and surgically supplements its attention/FFN subspaces with donor-model patterns so the base's multimodal routing stays intact while visible stylistic character moves in. That framing makes it useful as a deliberate “character graft” experiment rather than a wholesale retrain.
Key Capabilities
- Targeted grafting: Donor patterns were injected primarily into the two-token_refiner front blocks (blocks 0–1) and selected main-block positions, so text conditioning and downstream generation receive the stylistic shifts without replacing H3’s modality interfaces.
- Cross-model surgery: LTX-2.3 contributed attn_v-style content to carry motion/visual output character; Wan 2.2 supplied MLP fc1 patterns aligned to H3’s token_refiner inner dims; Krea 2 contributed attn_q (and limited k/v/input-side MLP) for finer spatial detail. q/k/v were temporarily unfused to allow per-projection control and then re-fused for inference.
- Orthogonal projection math: Additive, orthogonalized inserts bias H3 in directions it previously underused, supplementing rather than overriding preexisting capabilities and helping preserve temporal coherence and audio behavior where possible.
- Practical trade-offs: The merge preserves prompting behavior and most guardrails but produces measurable, limited audio degradation due to H3’s unified audio–video attention. License implications: MiniMax H3’s community license applies and donor-model community terms apply to the contributed character portions.
Who it’s for — and trade-offs
Great fit if you want an experimental pathway to transfer recognizable motion/visual styling from other image/video diffusion models into H3 without full re-training: researchers exploring model surgery, artists testing novel t2v aesthetics, and engineers prototyping cross-model merges. Look elsewhere if you need a production-hardened release with guaranteed zero audio change or formal provenance separation of learned components; this is an experiment with controlled additive perturbations, not a certified retrain.
