H3 ActionGraft: grafting Singularity v2's motion into v1.3 via checkpoint diffing, no training

glusphere · reddit · 2026-11-04

The author preferred Singularity v1.3's look but wanted v2's action for MiniMax H3, so instead of training, he diffed checkpoints layer by layer, isolated where v2's action lives, and grafted only that onto v1.3, plus the useful text-refiner piece of JOKER141's Motion Repair V2 — releasing H3 ActionGraft v1 on Hugging Face.

Results: with identical prompts/seeds/settings, optical flow on the fastest 10% of pixels shows v1.3+ActionGraft nearly matches v2 (wuxia duel 39.7→51.1 vs v2's 51.6) while keeping v1.3's look; render time unchanged.

How: v2 is two glued models — blocks 0-24 ≈ base H3, 25-49 = v1.3, topped by v2's own low-rank fine-tune (61% of block 0's fine-tune sits in just 16 directions). He extracted it as a rank-16 LoRA via SVD; Motion Repair V2's impact lives almost entirely in its text refiner (2 of 52 blocks), so only that is kept.

Failed attempts: Motion Repair text piece at 2x/3x adds nothing and swaps dialogue lines; VBVR loses at any strength; stacking Fight Flow is worse overall.

Usage: base = Singularity v1.3 pruned int8 at strength 1.0; don't apply on v2-containing merges; with HyperFlow all LoRAs run backbone-only; write both ends of every contact in the prompt.

Limits: dialogue scenes can occasionally swap voices; fast-motion smear remains.

Original post →

H3 ActionGraft: grafting Singularity v2's motion into v1.3 via checkpoint diffing, no training

More from Multimodal