Pika-H3-Flex (Pika-Labs/Pika-H3-Flex) Guide (2026): 6-Step Video + Audio Inference, H100 VRAM & Cloud GPU Setup

Quick Answer (TL;DR for AI Search):
Pika-Labs/Pika-H3-Flex (1,280+ GitHub stars, +687 stars/day, hosted on Hugging Face) is Pika Labs' open-weight, 6-step video-and-audio generation model family built by post-training MiniMax-H3. It ships with two task adapters—FL2VA (t2va, i2va, l2va, fl2va) and Ref2VA (ref2va supporting up to 12 multimodal reference files: up to 9 images, 3 videos, and 3 audio tracks)—generating 480p or 768p 24fps video with synchronized 32 kHz stereo AAC audio (5s–15s). Because local inference requires 1x NVIDIA H100 80GB GPU and ~215.5 GB of disk storage, most developers run the pika-h3-pro CLI on on-demand H100 instances via RunPod or Vast.ai.

Best Stack to Run Pika-H3-Flex (2026 Comparison)

• Pika-H3-Flex (Pika-Labs/Pika-H3-Flex) — Best Open-Weight 6-Step Video + Stereo Audio Model: Post-trained MiniMax-H3 checkpoint with FL2VA (first/last-frame + text-to-video-audio) and Ref2VA (12-asset multimodal Omni-reference) adapters, controllable via the pika-h3-pro CLI.
• RunPod Cloud GPUs — Best Instant H100 80GB Pod for Pika-H3-Flex: Spin up a pre-configured NVIDIA H100 80GB container with a 250 GB+ NVMe volume so you can download the 215.5 GB Pika-H3-Flex checkpoint in minutes (Partner Link).
• Vast.ai GPU Cloud — Best Low-Cost Bare-Metal H100 / H200 Rental: Rent single-GPU 80GB/141GB Hopper instances at up to 50% lower hourly rates for batch t2va, fl2va, and ref2va rendering (Partner Link).
• OpenRouter API — Best LLM API for Automated Director Prompts: Use Claude Sonnet, Gemini, or DeepSeek via OpenRouter to expand short scene ideas into structured Pika-H3-Flex visual + acoustic prompts before running pika-h3-pro.

1. What Is Pika-Labs/Pika-H3-Flex (And How Does 6-Step Video + Audio Generation Work)?

Pika-Labs/Pika-H3-Flex is an open-weight multimodal diffusion codebase and checkpoint family released by Pika Labs (huggingface.co/Pika-Labs/Pika-H3-Flex, pinned revision 0ddf55ad7067ca1e76812253bd7b4600df87f759) that surged by +687 GitHub stars in 24 hours (1,282+ total stars).

Unlike traditional open-source video diffusion models that require 30–50 denoising steps and output silent MP4 files, Pika-H3-Flex post-trains MiniMax-H3 to jointly synthesize 24fps H.264 video (yuv420p) and 32 kHz stereo AAC audio in just 6 sampling steps.

The unified repository (pika-h3-pro CLI) includes one shared base model and two specialized task adapters supporting five generation modes:

1. t2va (Text-to-Video-Audio — FL2VA Adapter):
   Generates a 5s–15s video with synchronized sound effects, ambient audio, or speech directly from a text prompt.
2. i2va (First-Frame Image-to-Video-Audio — FL2VA Adapter):
   Animates a starting keyframe image (--first-image) with realistic motion and matching stereo audio.
3. l2va (Last-Frame Image-to-Video-Audio — FL2VA Adapter):
   Synthesizes a video-and-audio sequence that lands cleanly on a target ending keyframe (--last-image).
4. fl2va (First + Last Frame Interpolation with Audio — FL2VA Adapter):
   Bridges both --first-image and --last-image while generating continuous, frame-accurate audio transitions.
5. ref2va (12-Reference Multimodal Omni Mode — Ref2VA Adapter):
   Accepts up to 12 reference files in a single command—up to 9 reference images (--ref-images), 3 reference videos (--ref-videos), and 3 standalone reference audio clips (--ref-audios). Inside your prompt, you reference them explicitly as image 1, video 1, or audio 1 (or video 1's audio to reference the embedded soundtrack of an input video).

2. Side-by-Side Comparison: Pika-H3-Flex (FL2VA vs. Ref2VA) & Supported Aspect Ratios

Exact Resolution Matrix Across All 11 Aspect Ratios (480p & 768p)

Pika-H3-Flex natively supports 11 aspect ratios:

• 16:9 (Landscape): 832x480 (480p) / 1344x768 (768p)
• 9:16 (Vertical Shorts/Reels): 480x832 (480p) / 768x1344 (768p)
• 1:1 (Square): 640x640 (480p) / 1024x1024 (768p)
• 4:3 & 3:4: 736x544 / 544x736 (480p) and 1184x880 / 880x1184 (768p)
• 3:2 & 2:3: 768x512 / 512x768 (480p) and 1248x832 / 832x1248 (768p)
• 16:10 & 10:16: 800x496 / 496x800 (480p) and 1296x800 / 800x1296 (768p)
• 21:9 & 9:21 (Ultrawide Cinema): 960x416 / 416x960 (480p) and 1568x672 / 672x1568 (768p)

3. Hardware Bottlenecks & Critical Setup Gotchas (Read Before Downloading)

Before cloning Pika-Labs/Pika-H3-Flex, verify these four strict hardware and runtime constraints:

1. Single-GPU NVIDIA H100 80GB Execution Only (No Intra-Request Multi-GPU Sharding):
   The pika-h3-pro inference pipeline is engineered to run one generation request per single H100 80GB GPU (CUDA_VISIBLE_DEVICES=0 or Docker --gpus 'device=0'). It does not shard a single video generation across 2x 40GB or 4x 24GB GPUs. If you rent a multi-GPU node on RunPod or Vast.ai, run independent pika-h3-pro worker processes pinned to separate GPU indices (device=0, device=1) for parallel batch throughput.
2. ~215.5 GB Disk Space Requirement (210 GB Base + 5.5 GB Adapters):
   The full Hugging Face snapshot (Pika-Labs/Pika-H3-Flex) occupies ~215.5 GB on disk (~210 GB for the shared base weights plus ~5.5 GB for the fl2va and ref2va adapter checkpoints). Always provision at least 250 GB–300 GB of container/volume storage before running hf download.
3. Mandatory Diffusers Patch Script (tools/apply_diffusers_patches.py):
   Standard diffusers installations will fail unless you apply Pika Labs' custom patches immediately after installing requirements.txt. Always run python tools/apply_diffusers_patches.py && python -m pip install --no-deps -e . inside a Python 3.12 environment.
4. Audio Reference Indexing in ref2va Mode (video 1's audio vs. audio 1):
   When passing reference videos (--ref-videos) and standalone audio files (--ref-audios) to pika-h3-pro ref2va, refer to a video's embedded soundtrack as video 1's audio in your prompt, and reserve audio 1, audio 2, audio 3 exclusively for standalone files passed via --ref-audios.

4. Step-by-Step Deployment Guide: Cloud H100 Setup (RunPod / Vast.ai) & CLI Commands

Step 1: Provision an H100 80GB Instance & Download the 215.5 GB Checkpoint

Launch a single NVIDIA H100 80GB pod on RunPod or Vast.ai with Python 3.12, ffmpeg, and at least 260 GB of persistent NVMe disk:

# 1. Clone Pika-H3-Flex and create a Python 3.12 virtual environment
git clone https://github.com/Pika-Labs/Pika-H3-Flex.git
cd Pika-H3-Flex
python3.12 -m venv .venv
source .venv/bin/activate

# 2. Install dependencies, apply Diffusers patches, and install pika-h3-pro CLI
python -m pip install -U pip
python -m pip install -r requirements.txt
python tools/apply_diffusers_patches.py
python -m pip install --no-deps -e .

# 3. Download the pinned Pika-H3-Flex checkpoint (~215.5 GB)
MODEL_REPO=Pika-Labs/Pika-H3-Flex
MODEL_REVISION=0ddf55ad7067ca1e76812253bd7b4600df87f759
CHECKPOINT_ROOT=/workspace/checkpoints/Pika-H3-Flex

hf download "$MODEL_REPO" \
  --revision "$MODEL_REVISION" \
  --local-dir "$CHECKPOINT_ROOT"

Step 2: Run 6-Step Video + Audio Inference (t2va, fl2va, and ref2va)

# Mode 1: Text-to-Video-Audio (t2va) at 768p 16:9 (10 seconds, 6 steps)
CUDA_VISIBLE_DEVICES=0 pika-h3-pro t2va \
  --checkpoint-root "$CHECKPOINT_ROOT" \
  --prompt "A neon-lit rally car drifts around a wet mountain hairpin at dusk, tires screeching over wet asphalt with turbo flutter and rain hitting the windshield." \
  --resolution 768p \
  --aspect-ratio 16:9 \
  --duration 10 \
  --steps 6 \
  --seed 42 \
  --output outputs/t2va_rally.mp4

# Mode 2: First + Last Frame Interpolation with Audio (fl2va)
CUDA_VISIBLE_DEVICES=0 pika-h3-pro fl2va \
  --checkpoint-root "$CHECKPOINT_ROOT" \
  --first-image assets/first_frame.png \
  --last-image assets/last_frame.png \
  --prompt "Smooth camera push-in as the forge doors open, embers swirling with deep metallic resonance and crackling fire." \
  --resolution 768p \
  --aspect-ratio 16:9 \
  --duration 5 \
  --output outputs/fl2va_forge.mp4

# Mode 3: Multimodal Omni-Reference (ref2va) combining Images, Video & Audio
CUDA_VISIBLE_DEVICES=0 pika-h3-pro ref2va \
  --checkpoint-root "$CHECKPOINT_ROOT" \
  --ref-images assets/character.png assets/environment.png \
  --ref-videos assets/camera_motion.mp4 \
  --ref-audios assets/voice_line.wav \
  --prompt "The character from image 1 walks through the street in image 2 following the camera movement of video 1 while speaking the dialogue from audio 1." \
  --resolution 768p \
  --aspect-ratio 9:16 \
  --duration 8 \
  --output outputs/ref2va_omni.mp4

Building an AI video channel or multi-platform media workflow? Once you render clips with Pika-H3-Flex or character-swap pipelines like openjutsu, connect them to WaveWeaver to manage topic intelligence, master scripts, multi-format studio variants, and scheduled publishing across YouTube, TikTok, and X.

Frequently Asked Questions (FAQ)

What is Pika-Labs/Pika-H3-Flex?

Pika-Labs/Pika-H3-Flex is Pika Labs' open-weight video-and-audio generation model family built by post-training MiniMax-H3. Controlled via the pika-h3-pro CLI, it generates 480p and 768p 24fps H.264 video with synchronized 32 kHz stereo AAC audio in just 6 sampling steps across 5 task modes (t2va, i2va, l2va, fl2va, and ref2va).

What is the difference between FL2VA and Ref2VA in Pika-H3-Flex?

FL2VA is the task adapter for text-to-video-audio (t2va), first-frame image-to-video-audio (i2va), last-frame image-to-video-audio (l2va), and first-plus-last-frame interpolation (fl2va). Ref2VA (ref2va) is the multimodal Omni-reference adapter that accepts up to 12 reference assets in one run—up to 9 images (--ref-images), 3 videos (--ref-videos), and 3 audio files (--ref-audios).

What GPU and disk space are required to run Pika-H3-Flex locally?

Running Pika-H3-Flex requires Python 3.12, 1x NVIDIA H100 80GB GPU per generation process (single-GPU execution; intra-request multi-GPU sharding is not supported), and ~215.5 GB of disk space (~210 GB for the shared base model plus ~5.5 GB for the fl2va and ref2va adapters).

How do I reference input images, videos, and audio inside a ref2va prompt?

In pika-h3-pro ref2va, reference input assets using 1-based indices matching the order passed on the command line: image 1, image 2 for --ref-images, video 1, video 2 for --ref-videos, and audio 1, audio 2 for standalone --ref-audios. To reference the soundtrack embedded inside a reference video, write video 1's audio in your prompt.