Pika H3 Flex in ComfyUI: Local Setup and Troubleshooting

You can set up the released Pika H3 Flex adapters in ComfyUI using matching MiniMax H3 components and the official workflows. This is a documentation-based setup guide, not a successful local deployment report. This independent website currently offers listed models such as MiniMax H3, not Flex itself.

Sources checked November 9, 2026. We have not installed these weights, run a GPU generation or measured compatibility, minimum VRAM, disk use, speed or output quality. The commands and settings below follow official instructions; they have not been executed here.

1. Check the environment and license first

Start with the official Flex model card and its linked hardware documentation. Verify the exact checkpoint format, backend and GPU support for your environment before renting hardware or downloading weights. General ComfyUI system support does not prove that a particular Flex workflow or its quantization kernels work on every supported device.

Question	Status in this guide	What you need to check
Required components and template settings	Checked against official documentation	Match the current template and its download links
Minimum VRAM, RAM and free disk space	Not measured by this site	Exact files, precision, offloading and decoding settings
NVIDIA, AMD, Intel or Apple compatibility	Not tested by this site	Backend, dependencies and operations used by this workflow
Generation time and quality	Not benchmarked by this site	Measure a completed run on the exact configuration

Read the privacy and commercial-use guide before using client assets or outputs. Downloadable weights do not grant unrestricted commercial or worldwide use. Local compute also has hardware, storage and operating costs; this site’s Credits do not pay for it.

2. Install or update ComfyUI

For a packaged installation, choose the appropriate route from the official installation page. On Windows, the Portable guide describes extracting the matching package and starting its launcher. That installs ComfyUI; it is not a Flex compatibility certification.

For a manual installation, follow the official manual guide. Create and activate an isolated Python environment, clone the official repository and install the PyTorch build appropriate to your hardware as instructed there. Then, in that environment:

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
python -m pip install -r requirements.txt
python main.py

Open the local address printed by the server. These commands assume Git, the isolated environment and the matching PyTorch dependencies are already available. They are not a universal GPU setup command.

For an existing manual installation, back up workflows and local changes, stop the server, activate its environment, and run inside its ComfyUI directory:

git pull
python -m pip install -r requirements.txt
python main.py

Use the package’s own updater for Portable or Desktop installations. The official update guide covers those differences. Do not update an unrelated system Python and assume the active ComfyUI runtime changed.

3. Pick one official workflow

In Template Library → Video, search for Pika H3 Flex. The ComfyUI Flex tutorial links three workflow JSON files:

Workflow	Input	Published minimum ComfyUI version
Text to Video	Prompt	0.43.2
Reference to Video	Reference images plus prompt	0.30.0
First and Last Frame / FL2VA	Opening and ending images plus prompt	0.30.0

These are documentation requirements checked on the date above, not versions tested by us. If a template is missing, update the actual runtime and template dependencies using the official instructions, restart and check again. You can also download its JSON from that tutorial and load it in ComfyUI. Do not replace a missing official template with an unreviewed workflow and treat the result as equivalent.

4. Download the matching model set

Use the download links in the selected official template’s tutorial. Download the base and adapter for that mode plus the shared encoder and VAEs. You do not need both mode-specific pairs just to try one workflow.

Mode	Diffusion model stem	Flex LoRA stem
Text / first and last frame	minimax_h3_fl2va_pruned_int8_convrot	pika_h3_flex_fl2va_v1_bf16
Reference	minimax_h3_ref2va_pruned_int8_convrot	pika_h3_flex_ref2va_v1_bf16

The published layout uses the following paths; install only the pair you selected and the shared files:

ComfyUI/models/
  diffusion_models/
    minimax_h3_fl2va_pruned_int8_convrot.safetensors
    minimax_h3_ref2va_pruned_int8_convrot.safetensors
  loras/
    pika_h3_flex_fl2va_v1_bf16.safetensors
    pika_h3_flex_ref2va_v1_bf16.safetensors
  text_encoders/
    qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
  vae/
    minimax_h3_video_vae_int8_convrot.safetensors
    minimax_h3_audio_vae_fp32.safetensors

Confirm complete downloads, file names and the loader selections after restarting ComfyUI. An external model directory needs a configured search path; see the manual installation guide’s extra_model_paths.yaml instructions. File size and memory use must be checked for your actual download and runtime. This guide does not recommend a minimum GPU or guarantee that a quantized file fits a particular card.

5. Keep the released sampling settings

The official template specifies a six-step schedule with lcm, ManualSigmas, and CFG 1.0 through BasicGuider, with adapter strength 1.0. Preserve the template’s sigma values and match FL2VA with FL2VA or Ref2VA with Ref2VA. Do not add another Turbo or Lightning acceleration LoRA. These settings are taken from the official tutorial, not tuned in a run by this site.

Start with a simple prompt and licensed, non-sensitive inputs. Keep the template defaults for the first diagnostic attempt; change one input or setting at a time. The prompt guide explains reference roles and sound descriptions.

When a run completes, verify that the output opens, has the intended duration and framing, and contains the expected audio. The model card describes local native 480p/768p with mixed stereo audio; this does not include all of Pika’s hosted high-resolution finishing or separate audio stems. Save your workflow, versions, seed if exposed, and actual settings with the result.

6. Diagnose by the failing stage

Symptom	First checks
Missing template or unknown node	Runtime version, restart, template dependencies and official workflow revision; avoid installing every suggested custom node blindly
Model absent from a loader	Correct folder/search path, complete download, exact filename and selected mode
Tensor shape or weight mismatch	Base/adapter pair, encoder, VAEs and precision expected by the template; preserve the original error
Out of memory	Identify loading, sampling or decoding stage; close competing processes and inspect documented memory options. Change one supported setting at a time; no setting here guarantees success
Broken motion or unusable output	Restore the unmodified template, adapter strength and sampling schedule before tuning the prompt
Missing sound or export error	Audio VAE selection, audio connections, final save node and its encoding error; compare with the official JSON

Record the ComfyUI version, OS, GPU/backend, driver and PyTorch versions, workflow JSON, model filenames, settings and first error. Remove private prompts, media URLs and credentials before sharing logs. For this website’s browser/API tasks, use the separate video troubleshooting guide.

7. Local inference and privacy

Inspect every node before introducing confidential material. A workflow can include cloud API nodes, remote URLs, custom code, extensions or uploads even when its interface opens on localhost. Model downloads also require network access. A local page address alone does not establish that inference is offline.

If offline handling is required, first obtain the needed dependencies and permitted files, audit the workflow and outbound connections, then verify it with non-sensitive inputs under the intended network restrictions. Protect local outputs, logs and workflow metadata too. See privacy, deletion and commercial-use boundaries for the cloud-service differences and applicable licenses.