Pika H3 Flex

What is Pika H3 Flex?

Pika H3 Flex is MiniMax H3 with additional training by Pika.

Pika also builds its own video generation models, and the training used its own dataset.

On top of better quality, step distillation cuts it down to 6 steps, so generation feels very smooth.

Basic usage is the same as MiniMax H3. See that page for how to write prompts and which nodes to use.

Recommended Settings

• Resolution
  • 768p, about 1 MP
  • Width and height must be multiples of 32
• FPS
  • 24 FPS
• Frames
  • 17n + 5
• Video length
  • 5–15 seconds

Model Download

LoRA

• loras
  • pika_h3_flex_fl2va_v1_bf16.safetensors (1.96 GB)
  • pika_h3_flex_ref2va_v1_bf16.safetensors (1.96 GB)

📂ComfyUI/
└── 📂models/
    └── 📂loras/
        ├── pika_h3_flex_fl2va_v1_bf16.safetensors
        └── pika_h3_flex_ref2va_v1_bf16.safetensors

Base Models

• checkpoints
  • sam3.1_multiplex_fp16.safetensors (1.75 GB)
• diffusion_models
  • minimax_h3_fl2va_pruned_int8_convrot.safetensors (21 GB)
  • minimax_h3_ref2va_pruned_int8_convrot.safetensors (21 GB)
• text_encoders
  • qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7 GB)
• vae
  • minimax_h3_audio_vae_fp32.safetensors (605 MB)
  • minimax_h3_video_vae_int8_convrot.safetensors (2.81 GB)

📂ComfyUI/
└── 📂models/
    ├── 📂checkpoints/
    │   └── sam3.1_multiplex_fp16.safetensors
    ├── 📂diffusion_models/
    │   ├── minimax_h3_fl2va_pruned_int8_convrot.safetensors
    │   └── minimax_h3_ref2va_pruned_int8_convrot.safetensors
    ├── 📂text_encoders/
    │   └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    └── 📂vae/
        ├── minimax_h3_audio_vae_fp32.safetensors
        └── minimax_h3_video_vae_int8_convrot.safetensors

text2video / T2VA

Model: fl2va

• Load LoRA
  • Apply pika_h3_flex_fl2va_v1_bf16.safetensors at strength 1.0
• SamplerCustomAdvanced + ManualSigmas
  • Use the 6-step sigmas Pika specifies
  • 1.0, 0.972973, 0.923077, 0.8, 0.571429, 0.333333, 0.0
  • The sampler is lcm

image2video / I2VA

Model: fl2va

Outputs

First / Last Frame to Video

Model: fl2va

Outputs

Reference Generation

Model: ref2va

• Load LoRA
  • Apply pika_h3_flex_ref2va_v1_bf16.safetensors at strength 1.0

Video Editing

Model: ref2va

Spatial Inpainting

Model: ref2va

Outputs

References

• Pika H3 Flex
• Pika H3 Flex ComfyUI Workflows