Running Pika H3 Flex in ComfyUI
Technical guide for deploying Pika H3 Flex via ComfyUI: workflow selection, model file directories, normalized 6-step sampler sigmas, and legal licensing terms.
The file lists, directory paths, node configurations, and sampling values documented below are compiled directly from the published documentation of Pika Labs, Comfy-Org, and MiniMax AI. This guide documents published specifications and community repository instructions; we have not executed an independent local GPU verification benchmark.
Workflow Selection & ComfyUI Version Requirements
Official ready-to-run ComfyUI templates are pre-packaged in up-to-date ComfyUI installations under Template Library > Video. Version requirements documented by Comfy-Org include:
1. Text-to-Video (T2V)
LoRA Adapter: pika_h3_flex_fl2va_v1_bf16.safetensors
Version Requirement: ComfyUI v0.43.2+ (or current ComfyUI Desktop app release).
When to choose: Generating video sequences entirely from narrative text descriptions without starting images or reference clips.
2. First & Last Frame (FL2VA)
LoRA Adapter: pika_h3_flex_fl2va_v1_bf16.safetensors
Version Requirement: ComfyUI v0.30.0+.
When to choose: Providing anchor keyframes (first frame, last frame, or both) to control opening and closing compositions while generating smooth visual and audio transitions between them.
3. Reference-to-Video (Ref2VA)
LoRA Adapter: pika_h3_flex_ref2va_v1_bf16.safetensors
Version Requirement: ComfyUI v0.30.0+.
When to choose: Conditioning sequences on up to 12 multimodal references (up to 9 images, 3 videos, 3 audio clips) to maintain character identity, styling, or motion dynamics.
Official resources:
• Official Tutorial: docs.comfy.org/tutorials/video/pika/pika-h3-flex (accessed Oct 30, 2026)
• Model Repository: Comfy-Org/MiniMax-H3 on Hugging Face (Discussion / PR #70, accessed Oct 30, 2026)
Required Model Files & Target Directories
To run Pika H3 Flex within ComfyUI, base MiniMax H3 weights and companion models must be positioned within your ComfyUI root directory structure:
| Destination Directory | File Name | Format / Precision | Functional Role |
|---|---|---|---|
models/diffusion_models/ |
minimax_h3_fl2va_pruned_int8_convrot.safetensors |
INT8 quantized base | Base diffusion transformer for Text-to-Video and First/Last Frame |
minimax_h3_ref2va_pruned_int8_convrot.safetensors |
INT8 quantized base | Base diffusion transformer for Reference-to-Video | |
models/loras/ |
pika_h3_flex_fl2va_v1_bf16.safetensors |
BF16 LoRA adapter | Pika 6-step distillation adapter for FL2VA workflows |
pika_h3_flex_ref2va_v1_bf16.safetensors |
BF16 LoRA adapter | Pika 6-step distillation adapter for Ref2VA workflows | |
models/vae/ |
minimax_h3_video_vae_int8_convrot.safetensors |
INT8 quantized | Encodes and decodes video latent frames |
minimax_h3_audio_vae_fp32.safetensors |
FP32 audio VAE | Decodes audio latents into synchronized stereo audio | |
models/text_encoders/ |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
NVFP4 / AWQ quantized | Vision-language text encoder for prompt conditioning |
Pika CLI Scheduler Knots vs. ComfyUI ManualSigmas
A frequent point of confusion in community setups is the distinction between Pika's command-line interface (CLI) scheduler knots and ComfyUI's normalized sigma schedule:
Pika CLI Scheduler Knots (Integer Timesteps)
In Pika's native diffusers/CLI pipeline, the distillation schedule is represented as unnormalized integer timestep knots spanning from 1000 down to 0:
1000, 750, 500, 250, 100, 40, 0
These values define seven boundary points that carve out the six discrete denoising intervals in the baseline pipeline.
ComfyUI ManualSigmas (Normalized Floats)
ComfyUI expects normalized floating-point sigmas between 1.0 and 0.0. As documented in Comfy-Org discussion #70, the published sequence for the ManualSigmas node is:
1.0, 0.972973, 0.923077, 0.8, 0.571429, 0.333333, 0.0
Entering raw integer timesteps (like 1000 or 750) into ComfyUI's ManualSigmas input will result in calculation errors or visual distortion.
Verified Six-Step Sampler Node Pipeline
In ComfyUI, the sampling branch is configured as follows:
| Node Component | Configured Parameter | Requirement Notes |
|---|---|---|
LoraLoaderModelOnly |
strength_model: 1.0 |
Loads the corresponding Pika LoRA onto the base diffusion model. Strength must remain at 1.0. |
BasicGuider |
cfg: 1.0 |
Classifier-Free Guidance must remain fixed at 1.0. Values greater than 1.0 cause severe contrast clipping in distilled models. |
KSamplerSelect |
sampler_name: lcm |
Selects the Latent Consistency Model solver matching the distilled flow trajectory. |
ManualSigmas |
sigmas: 1.0, 0.972973, 0.923077, 0.8, 0.571429, 0.333333, 0.0 |
Exact 6-step normalized sigma array documented in Comfy-Org PR #70. |
SamplerCustomAdvanced |
Combines Guider, Sampler, Sigmas & Latents | Executes the denoising pass and outputs audiovisual latents to the VAE decoders. |
Storage Footprint & Hardware Considerations
Storage: The Pika CLI repository documents approximately 210 GB of files when downloading the full base repository across both task families. When using ComfyUI, creators typically download only the pruned, INT8-quantized weights and NVFP4 text encoder required for their specific task (e.g. FL2VA only), which requires significantly less disk space.
GPU Hardware: The base architecture comprises a 33B parameter diffusion transformer alongside a 32B vision-language text encoder. Memory requirements depend heavily on quantization level (e.g. INT8/FP4 vs BF16) and whether CPU offloading or layer swapping is enabled in ComfyUI. Neither Pika Labs nor Comfy-Org has published a verified consumer GPU minimum. Do not assume standard consumer GPUs can execute these models locally without independent testing on your specific system.
Licensing & Legal Constraints
Review the specific legal terms governing the code and model weights before deploying locally:
Pika Code: Apache 2.0
Pika’s Python inference code, Diffusers pipeline wrappers, and utility scripts in the Pika-Labs/Pika-H3-Flex GitHub repository are released under the Apache 2.0 license.
This allows code inspection, modification, and redistribution. However, this code license explicitly does not cover or grant rights to the underlying model weights.
Model Weights: MiniMax H3 Community License
The model weights are subject to the MiniMax H3 Community License Agreement (dated August 2, 2026):
- Territorial Restrictions: Under the community license, the Applicable Territory is worldwide but strictly excludes the European Union, the United Kingdom, the Republic of Korea, and the United States of America. Using the weights locally within these excluded territories requires separate written authorization or a commercial license from MiniMax.
- Yearly Revenue Condition: In permitted territories, commercial use is permitted under the community license unless the relevant commercial products and services generate more than US $20 million in yearly revenue, in which case licensees must obtain separate prior written authorization from MiniMax.