Pika H3

How to Use Pika H3: a 5-minute walkthrough

Pika H3 is the short name for Pika H3 Flex, Pika Labs' post-trained MiniMax H3 video model. This page walks through the fastest path to a watchable clip, hosted or local. Updated November 2026.

Hosted path (no install)

  1. Open the Pika H3 generator.
  2. Pick a mode: FL2VA for text or frame pairs, Ref2VA for references.
  3. Write a short, concrete prompt — one subject, one camera move.
  4. Leave steps at 6. Pick duration (10s is a good default) and resolution (1080p).
  5. Fix the seed while iterating. Fan out across seeds once wording is right.
  6. Export the MP4.

Local path (ComfyUI)

  1. Install the latest ComfyUI nightly (docs.comfy.org).
  2. Pull FL2VA and Ref2VA safetensors from huggingface.co/Pika-Labs/Pika-H3-Flex into models/diffusion_models/.
  3. Load the official workflow JSON. Queue a prompt at steps 6.

Common first-run mistakes

  • Writing a long, adjective-heavy prompt. Shorter wins.
  • Random seed every run. Fix it while iterating.
  • Mixing two camera moves in one prompt. The model picks one and wobbles.
  • Chasing 4K before 1080p looks right.

FAQ

Do I need a GPU to use Pika H3?

Only if you run it locally in ComfyUI. The hosted generator above runs on remote hardware, so any browser works.

How long does one generation take?

About seven seconds for a fifteen-second clip at 1080p on a modern 24 GB GPU. The hosted generator adds queue time; local generation is bound only by your hardware.

Can I change the duration after I pick it?

Not within the same generation. Change the duration slider and re-render; the sampler always runs end-to-end.

Related