sekkit/Pika-H3-Flex-Turbo-LoRA

Pika H3 Flex Turbo LoRA — Pika's motion, one file, at every speed: 2-step draft → 4-step fast → 8-step max quality

Pika H3 Flex fused with the best few-step distills for MiniMax-H3 into single standard LoRAs (exact rank-concatenation — no custom nodes, no model patches). Drop one file onto any stock H3 setup — ComfyUI or Diffusers, bf16 / int8 / GGUF — or onto quality merges like Singularity, where it gets even better. Video with synchronized audio.

⭐ The combo: 4-step Turbo LoRA on Singularity vs Pika official on stock H3

Left: Pika official Flex adapter on stock H3, 6 steps. Right: this 4-step LoRA on Singularity v2, 4 steps. Five prompts, same seed, 848×480 with stereo audio. Pika's motion + Singularity's quality tuning + turbo speed — in our tests the right side is both faster and better-looking than official. Uses the Ref2VA 4-step file (Singularity ships ref2va).

⚡ 2-step draft mode on Singularity — 5 fresh head-to-heads

Left: official 6-step. Right: the 2-step draft LoRA on Singularity — ballet leap, F1 cornering, diner kitchen, lighthouse storm, reef turtle. Flex × pDMD 2-NFE, ~14 s of sampling per 5-s clip (2 evaluations at ~7 s/step on H200). First test batch (5 more prompts): 4 of 5 held up well. Use 2-step to iterate framing and motion, then render finals at 4 or 8 steps — occasional scene failures happen, it's labeled draft for a reason.

Speed: Pika official (6 steps) vs this LoRA (4 steps) on stock H3

Same five prompts, stock int8 H3, seed 42: official 6-step (left) vs this LoRA 4-step (right).

Max quality on stock: official (6 steps) vs 8-step HD

The 8-step HD variant (style 0.6) for maximum detail on stock H3. Also transfers to Singularity: 8-step HD on Singularity vs official.

High-action stress test

Martial arts sparring + skateboard hill-bomb: official 6-step (left) vs 4-step turbo (right). Few-step models are weakest on fast motion — judge it here.

Benchmarks (measured — same base, seed, prompt, CFG 1, 848×480 × 121 frames + stereo audio)

	Steps	End-to-end per 5-s clip	vs official
H200 — Pika official adapter	6	60 s	1×
H200 — this LoRA, 4-step	4	40 s	1.5× faster
H200 — this LoRA, 2-step draft	2	~25 s (14 s sampling)	~2.4× faster
H200 — this LoRA, 8-step HD	8	70 s	0.86× — quality-max
A100 — Pika official adapter	6	109.3 s	1×
A100 — this LoRA, 4-step	4	66.0 s	1.66× faster
Base MiniMax-H3 (~40 steps)	~40	~500 s	~8–12× slower

Same per-step cost on identical hardware (~7 s/step H200, ~12.5 s/step A100) — the speedup comes entirely from needing fewer evaluations.

Quality: every file is a mathematically exact composition of its source adapters (per-layer relative error ≈ 1–3e-7 — it is the stacked adapters, as one file), so fidelity to the stacked reference is exact by construction; Flex style is baked at 0.5 (2/4-step) / 0.6 (8-step HD), the community-validated operating points for few-step H3. Judge the look in the side-by-sides above.

Lite variants: rank-32 SVD compression (exact QR-product truncation) retains 96.1% of the fused spectral energy at 0.31 GB — 12.6× smaller.

The lineup

File	Transformer	Steps	Size	Pick this for
pika_h3_flex_turbo_2step_draft_ref2va.safetensors	Ref2VA	2	3.9 GB	Draft/iterate — pDMD fusion, check framing & motion fast
pika_h3_flex_turbo_4step_fl2va.safetensors	FL2VA	4	3.9 GB	Speed — fastest production Pika Flex
pika_h3_flex_turbo_4step_ref2va.safetensors	Ref2VA	4	3.9 GB	Speed + refs · the Singularity combo file
pika_h3_flex_turbo_8step_hd_fl2va.safetensors	FL2VA	8	3.9 GB	Quality — max detail, style 0.6
pika_h3_flex_turbo_8step_hd_ref2va.safetensors	Ref2VA	8	3.9 GB	Quality, with multi-reference identity
pika_h3_flex_turbo_4step_fl2va_lite_r32.safetensors	FL2VA	4	0.31 GB	Tiny — 96% spectral energy
pika_h3_flex_turbo_4step_ref2va_lite_r32.safetensors	Ref2VA	4	~0.3 GB	Tiny, with multi-reference identity

Unlike patch-based or custom-node releases, these are plain LoRAs — they compose with your existing checkpoints, other (gentle) LoRAs, upscalers and workflows, in every H3-compatible loader.

Usage (ComfyUI)

1. Download the file for your transformer into ComfyUI/models/loras (Ref2VA files for Singularity / reference workflows, FL2VA for stock T2V/I2V).
2. Load your usual MiniMax-H3 workflow, add a LoraLoaderModelOnly, strength 1.0.
3. Sample: 2, 4, or 8 steps (match the file), CFG 1.0, euler/lcm/res_multistep, simple scheduler. Flow shift: video 6.0 (FL2VA) / 12.0 (Ref2VA profile), audio shift 3.0.

Tips

• The 4-step files are tuned to be run at exactly 4 steps; the 8-step HD files reward 6–8 steps; the 2-step file is draft-only — expect occasional scene failures and use it for iteration.
• Flex style is baked at 0.5 (2/4-step) / 0.6 (8-step) — the empirically optimal strength for stacking a style adapter on few-step-distilled H3 (community consensus range 0.4–0.8, lower end for fewer steps).
• The 4-step turbo half is lightx2v v1.2 768p for FL2VA (v1.3 has an acknowledged stacking regression); Ref2VA uses the field-proven ref2v v0.1; the 2-step half is pDMD 2-NFE (Apache 2.0).
• Audio degrades below ~6 steps (true of every H3 turbo) — worst at 2 steps. Render finals at 4+ if audio matters.
• Avoid stacking additional strong style/motion LoRAs at high strength on the 2/4-step files.

Why the quality holds: exact fusion, not an approximate merge

Each file is an exact rank-concatenation of the two adapters, per layer:

B_fused @ A_fused  ==  s·(B_flex @ A_flex) + 1.0·(B_turbo @ A_turbo)

Functionally identical to loading both LoRAs stacked — by construction, with zero merge interference. Naive additive merges collide in the output-side B matrix (arXiv 2604.16826); concatenation cannot. Verified numerically per layer before upload (worst relative error ~2.6e-7 through a bf16 round-trip), then verified end-to-end on GPU (the benchmarks and demos above).

Sources: Pika-Labs/Pika-H3-Flex (Pika / Mellis, Inc.), lightx2v/Minimax-h3-Turbo, pdmd2026/pdmd_2NFE_lora via Iwannapose's ComfyUI conversion (Apache 2.0). Base model: MiniMaxAI/MiniMax-H3. Demonstrated on WarmBloodAban/Minimax-h3_Singularity. Method background: PTD (arXiv 2610.11498), WaDi (arXiv 2603.08258). Reproduction scripts: scripts/.

License

MiniMax H3 Community License Agreement — Model Derivative of MiniMax H3 (Section III.2); the same license applies to your use and derivatives. See the base model license. Credit: MiniMax (base), Pika / Mellis, Inc. (Flex adapters), LightX2V (turbo distillation), pDMD authors (2-step distillation).