GitHub radar
Speed up MiniMax-H3 video with Pika's style — up to 2.4x faster
This is an add-on file for the MiniMax-H3 model that brings Pika's signature motion style and speeds up video rendering up to 2.4x — without reinstalling the model itself.
A LoRA is a small file that changes how an already-installed video model, MiniMax-H3, behaves, adding Pika's motion style to it.
When it helps
Useful if you already generate video with MiniMax-H3 on a powerful GPU (H200/A100-class, 24GB+ VRAM). Not worth it without MiniMax-H3 or on a regular laptop.
Pros
- Renders clips 1.5–2.4x faster, according to Pika's own numbers
- Video comes out with audio already synced to the picture
- A compressed version is 12x smaller at 96% of the quality
- Installs without reinstalling MiniMax-H3 — just drop in a file
Cons
- Speed was only measured on server-grade H200 and A100 cards
- MiniMax-H3 itself needs a GPU with 24GB+ VRAM even without this
- The fastest 2-step mode is a draft, not for final video
How to set it up — step by step
- 1Download the LoRA file from Hugging Face for the step count you need — 2, 4, or 8.
- 2Put the file into the loras folder of your ComfyUI setup (a video-workflow app).
- 3Add a "LoraLoaderModelOnly" step to your generation workflow.
- 4Send your agent the text from the block below.
- 5Run a render at 2 steps for a draft, or 4–8 steps for a final clip.
- 6Check the result — the audio should match the picture.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open this LoRA https://huggingface.co/Pika-Labs/Pika-H3-Flex-Turbo-LoRA, explain how to plug it into my MiniMax-H3 setup in ComfyUI, and show which file variant (2, 4, or 8 steps) to use for drafts versus final renders.
Needs a serious GPU — the authors only benchmark on datacenter-grade H200 and A100 cards, and the underlying MiniMax-H3 model itself needs a GPU with at least 24 GB of VRAM even without this add-on; a regular laptop will not handle it.
Open on Hugging Face▌ More finds