Grafting Krea2 into MiniMax H3: improving texture via component transplant
Key-Philosopher-9327 · reddit · 2026-08-24
The author attempted to graft Krea2's attention/MLP components into MiniMax H3 to fix texture issues ("plastic skin"), releasing custom ComfyUI nodes for testing.
Key Findings:
- Content Graft: Transplanting per-head Q/V/outproj and MLP. V weights are the safest and most effective lever (0.6–0.7), followed by MLP. MLP block range must stop at 44 to avoid lattice artifacts. Concepts land well, but character fidelity remains faint.
- QK-norm Transplant: Transplanting only the Q-norm gain (128 floats) improves rendering commitment, fixing big-object issues. Unlike Z-Image grafts, Krea2 requires the full 0–50 block range.
- Caveats: The orthogonalblend mode is numerically unstable; use directionblend. QK-norm grafting appears to break LoRAs trained on native H3.
The project relies on reverse engineering, building on methods by TenStrip and joeygambino.
More from Multimodal
- Converting 360° video to 3DGS using GauzillaPro and importing to UE — danbri · 2026-08-24
- Leonardo AI launches Wan 3.0 video generation model — aziz4ai · 2026-08-24
- MiniMax-H3 Fun Controlnet Union model released — physalisx · 2026-08-24
- Seedance 2.5 generates 30s native video, Magnific controls 3D motion and 4K — mhdfaran · 2026-08-24
- AI-generated 1974 wedding chat: Grandpa has six fingers? — mhdfaran · 2026-08-24
- Testing MiniMax Music 3: Cyberpunk 2077 song generation workflow — Ok-Entertainer-2991 · 2026-08-24