VideoToolMap
Local AI Videoالعربية

HUNYUANVIDEO 1.5 · LOCAL / OFFLINE · DOCUMENTATION-BASED · VERIFIED SEPTEMBER 13, 2026

How to Run HunyuanVideo 1.5 Locally: Requirements, Setup Paths & ComfyUI Workflow

HunyuanVideo 1.5 is an 8.3B video model with official Text-to-Video and Image-to-Video support, a documented 14GB VRAM minimum with model offloading, an official Linux/CUDA path, Diffusers support, and official ComfyUI workflow templates.

Technical research notice

This guide is based on Tencent Hunyuan and ComfyUI official documentation. Unless explicitly marked Hands-on Tested, VideoToolMap has not independently benchmarked this exact setup on owned hardware.

Documentation-basedNot independently hands-on tested by VideoToolMap

Quick verdict

Official minimum

14GB VRAM with model offloading enabled on an NVIDIA CUDA GPU.

Official software path

Linux, Python 3.10+, and a CUDA version compatible with PyTorch.

Generation modes

T2V and I2V with 480p/720p model families and super-resolution models.

ComfyUI

Official usage guide and workflow templates for 720p T2V and 720p I2V.

Hardware and software requirements

GPUNVIDIA GPU with CUDA supportOfficial source-code path is NVIDIA/CUDA-focused.
VRAM14GB minimum with model offloading enabledA documented run path, not a speed guarantee.
OSLinuxLinux is the official upstream software target.
Python3.10 or higherA clean environment reduces dependency conflicts.
CUDACompatible with installed PyTorchDriver/CUDA/PyTorch compatibility matters as much as raw VRAM.

Three official setup paths

1. Native source-code path

Clone the official repository, install requirements, download checkpoints, then use the provided inference script.

2. ComfyUI path

Tencent documents an official ComfyUI route using Comfy-Org repackaged checkpoints and native templates.

3. Diffusers path

The official README includes HunyuanVideo15Pipeline examples with CPU offload and VAE tiling.

ComfyUI workflow facts that matter

Use the latest ComfyUI and official workflow templates. With Tiled VAE decoding, Tencent says to set temporal_size to 4096 to avoid artifacts. For the 480p I2V step-distilled model, 8 or 12 steps are recommended; 4 steps is supported for more speed with some quality trade-off.

480p, 720p and super-resolution

Official configuration signals

Standard 480p T2V/I2V models use 50 steps; 480p I2V Step-Distilled recommends 8 or 12; 720p standard/CFG-distilled workflows use 50; 480→720 and 720→1080 super-resolution models are also listed.

Local is not always fully offline

Prompt rewriting is the key nuance. The official pipeline can use vLLM-compatible models or an external service. For a fully local workflow, disable rewriting or host the rewrite model locally.

OOM and memory notes

OOM

Even above 14GB, OOM can occur depending on workload. The README suggests PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:128 as one mitigation and exposes group-offloading controls when CPU memory is constrained.

LoRA and training

LoRA

HunyuanVideo 1.5 includes official training and LoRA support, but this guide focuses on inference. See the Local Character Consistency & LoRA guide.

License and commercial-use caution

HunyuanVideo 1.5 is governed by the Tencent Hunyuan Community License, not Apache-2.0. The current license includes territorial exclusions, acceptable-use restrictions and additional commercial terms, so VideoToolMap does not label it as unrestricted commercial use.

What we have not tested yet

Not independently hands-on tested by VideoToolMap

VideoToolMap has not yet measured generation time, peak VRAM, system RAM use, disk footprint, failure rate, output quality or stability on owned hardware.

Official sources

HunyuanVideo 1.5 — official repositoryOfficial ComfyUI Usage GuideComfyUI official workflow templatesTencent Hunyuan Community License

Last Verified: September 13, 2026 · Reference links do not imply affiliation, sponsorship or endorsement.