indhic-ai/H3-Singularity-ActionGraft

H3 Singularity ActionGraft

Current release: v1 (H3_ActionGraft_v1.safetensors). v2, aimed at stronger impacts, is in progress.

A LoRA for MiniMax H3 Singularity v1.3 (int8) that gives fights faster exchanges and harder impacts while keeping v1.3's look. It was made with no training at all. We dissected two existing models in weight space and kept only the parts that carry the action.

• Speed comes from Singularity v2's own fine-tune, pulled out of the v2 checkpoint as a rank-16 LoRA.
• Impacts and blocking come from Motion Repair V2's text-refiner piece (2 of its 52 blocks, 16 of its 416 tensors). Its side effects live in its other blocks, which are left out.

On our fight tests, the fastest-moving pixels move 40-55% faster than on v1.3. On our main test the ActionGraft arm lands six rapid hits in the middle of the fight where v1.3 lands two. A slow drama scene stays slow.

H3_ActionGraft_v1.safetensors: 155 MB, bf16, ComfyUI key format (diffusion_model.blocks.N.*, diffusion_model.token_refiner.*).

How to use it

• Base: Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8 (Singularity v1.3, pruned int8). Strength 1.0.
• Do not stack it on Singularity v2 or on merges that contain v2 (e.g. Singularity x 10Eros PRISM). They already have v2's fine-tune, so it would be counted twice.
• Tested with: Pika H3 flex ref2va LoRA (6 steps, euler/simple) in the MiniMax H3 Master Extender.
  • Pass 1 at 608x352, then a 1-step DMAD refine at 1280x720.
  • SLA 0.9 sparse attention, ActionLogic bridge V2, Sol VAE.
• HyperFlow caveat: the ComfyUI HyperFlow node turns its per-checkpoint curve fit off whenever ANY LoRA is on the model (curve.py: "requires an unmodified checkpoint"). So with HyperFlow, this LoRA (or any other) runs on the backbone only. To use it under HyperFlow, merge it into the checkpoint and build a curve fit for the merged file.
• Prompts: write both ends of every contact: the strike and what it does to the other body. For example, "her shield takes the club; she is driven back a step, boots sliding". The impact part of this LoRA strengthens contact that the prompt describes; it does not invent it.

Test videos

Every row uses the same prompt, references, seed and recipe. Only the model changes. Speed = mean flow of the fastest-moving 10% of pixels, in px/frame at 720p, between neighbouring frames (cuts skipped).

test	v1.3	v1.3 + ActionGraft v1	Singularity v2
Demon-hunter fight, 2 chained 15 s clips	28.4	42.4	45.4
Wuxia temple duel, 2 chained 15 s clips	39.7	51.1	51.6
Devi vs the Demon Horde, clip 2 (throat kill, brute charges)	29.3	45.2	43.2
Devi vs the Demon Horde, clip 7 (four-arm flurry vs the demon lord)	27.0	40.5	45.2
CHOODI s1e05 slow scene (all-pixel mean, see below)	0.32	0.55	-

Job time is unchanged: 158 vs 160 s, 181 vs 189 s, 101 vs 102 s, 100 vs 99 s, and 87 vs 92 s, v1.3 vs ActionGraft.

Demon-hunter fight (our standard test)

v1.3: v1.3 + ActionGraft v1: Singularity v2:

Wuxia temple duel

v1.3: v1.3 + ActionGraft v1: Singularity v2:

Devi vs the Demon Horde, clip 2

v1.3: v1.3 + ActionGraft v1: Singularity v2:

Devi vs the Demon Horde, clip 7

v1.3: v1.3 + ActionGraft v1: Singularity v2:

Slow drama: CHOODI s1e05 (a woman wakes to food left for her, goes quiet, answers a boy)

A check that the LoRA does not rush a still scene.

• Mean motion over all pixels went from 0.32 to 0.55 px/frame (at 480 px wide), still well under a pixel per frame.
• The scene gains a little life, and the dialogue stays with the right speakers.

v1.3: v1.3 + ActionGraft v1:

How it was made (no training)

1. Singularity v2 is two models glued together, plus a fine-tune

We diffed base MiniMax H3 ref2va, Singularity v1.3 and Singularity v2, all pruned int8, layer by layer.

• The int8 weights are stored Hadamard-rotated per 256-input group. The rotation is orthogonal and identical in every file, so differences can be measured on the stored weights directly.

	blocks 0-24	blocks 25-49
v1.3 vs base	3.3% change	3.3% change
v2 vs base	1.1-1.9%, unrelated to v1.3's change (cosine 0.03)	3.5%, the same direction as v1.3's (cosine 0.93)
what v2 did	undid v1.3 (cosine -0.9) and kept base	kept v1.3 and added a little

On top of that, v2 carries its own low-rank fine-tune.

• In block 0, 61% of v2's change sits in just 16 directions. Int8 rounding noise puts 2% there, and v1.3's own change 1%.
• That fine-tune matches none of the popular H3 action LoRAs: every cosine is below 0.005.

2. Extract v2's fine-tune as a LoRA

• A rank-16 SVD of (v2 - base) for blocks 0-24 and of (v2 - v1.3) for blocks 25-49, un-rotated to true weight space.
• Script: scripts/extract_lora.py.
• On v1.3, this alone gave v2's speed with v1.3's look. On our main test it was ranked above both v1.3 and v2.

3. Keep only Motion Repair V2's impacts

JOKER141's Motion Repair V2 adds impact and blocking, but it slows fights down, and on a dialogue scene it lip-synced a line onto the wrong character. We split it by block range (scripts/mr_split.py) and rendered each piece on top of step 2:

Motion Repair V2 piece	impacts (fight test)	fast-motion speed	dialogue scene
text refiner (token_refiner 0-1)	best	no slowdown	clean
blocks 14-33	second	slight	line mouthed by the wrong character
blocks 34-49	weak	slowest	line mouthed by the wrong character
blocks 0-13	weak	none	clean

The impacts live in the text refiner, the part that encodes the prompt. Our reading is that Motion Repair's training made contact words land harder. ActionGraft v1 = step 2 + that text-refiner piece, in one file (scripts/build_actiongraft.py).

Limitations

• Single seed per test. Every comparison is one seed, judged by eye plus the speed measurement. It is not a benchmark.
• Smear on fast motion is not fixed. It is reduced on some tests, but it is still there. Adding a "fast shutter" sentence to the prompt, or a second refine step, did not help.
• Physics is better but not perfect.
• Tested only on Singularity v1.3 int8 with the Pika 6-step LoRA at 720p.

Credits and license

• Singularity v1.3 / v2: WarmBloodAban/Minimax-h3_Singularity (Apache-2.0). The speed component is derived from v2's weights.
• Motion Repair V2: JOKER141/MiniMax-H3-General-Motion-Continuity-Repair (MiniMax H3 Community License). The impact component is its text-refiner tensors, unchanged.
• MiniMax H3: MiniMaxAI/MiniMax-H3. As a derivative of MiniMax H3, this LoRA is distributed under the MiniMax H3 Community License.
• Made by indhic-ai with dhee. Test films: our own (Devi vs the Demon Horde, CHOODI, demon-hunter and wuxia tests).