Local AI Video Generation in ComfyUI (2026): Wan 2.2, HunyuanVideo 1.5 and LTX-2.5

ComfyUI generates video as well as images, and in 2026 several open video models run on a single consumer GPU. Three families are natively supported: Wan 2.x from Alibaba, HunyuanVideo 1.5 from Tencent and LTX-2.5 from Lightricks. They differ a lot in memory needs, in output, and above all in licence. The upscale models you finish a render with have the same problem, and two of the most recommended are non-commercial. This guide covers what each needs, where the model files go, and the licence catches that matter before anything you make goes public.

Quick answer

• Lowest VRAM: Wan 2.2 TI2V-5B, which ComfyUI’s docs say “should fit well on 8GB vram with the ComfyUI native offloading”. Wan 2.1 T2V-1.3B is stated by Wan at 8.19GB.
• 24GB class: HunyuanVideo 1.5 (8.3B), which ComfyUI describes as delivering “flagship-quality video generation on consumer GPUs (24GB VRAM)”.
• Video with sound: LTX-2.5 (22B) generates synchronised audio and video, and ComfyUI ships three native workflows for it.
• Safest licence for commercial work: Wan, under Apache 2.0. LTX-2.5 needs a paid licence only if your entity turns over $10m or more. HunyuanVideo 1.5’s licence does not apply in the EU, the UK or South Korea.

The models

Image VRAM figures, like those in our image generation calculator, don’t carry over to video. A video model processes many frames at once, so memory depends heavily on resolution, length and offloading. The figures below are what each vendor or ComfyUI publishes, attributed. Nothing here is our own measurement.

Wan 2.2’s A14B models are a mixture of experts. Wan’s own description is that each expert is about 14B parameters, “resulting in a total of 27B parameters but only 14B active parameters per step”. As with language models, the active count sets the compute per step while the total still has to be held somewhere — the same distinction explained in our guide to tokens per second.

What the vendors say about speed

These are the vendors’ own claims, on their own hardware:

• Wan 2.1 T2V-1.3B: “It can generate a 5-second 480P video on an RTX 4090 in about 4 minutes (without optimization)”.
• Wan 2.2 TI2V-5B: “Without specific optimization, TI2V-5B can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU”.

Expect longer with more offloading, which is the trade you make to fit a smaller card.

Getting started in ComfyUI

1. Update ComfyUI first. Video support moves fast: ComfyUI shipped four releases between 15 September and 5 October 2026, ending at v0.39.0. ⚠️ ComfyUI’s own video pages warn that if nodes are missing when you load a workflow, the likely cause is that you are “not using the latest ComfyUI version (Nightly version)” — some video nodes land in nightly before they reach a stable release. New to ComfyUI? Our guide to running Stable Diffusion locally covers the install.
2. Open a template. Click the Templates icon in the sidebar, or use Workflow → Browse Workflow Templates. It lists the natively supported model workflows, video included.
3. Download the model files into the right folders (below). ComfyUI’s per-model pages list the exact file for each workflow.
4. Start small. Run a short clip at the lowest resolution the template offers, confirm it completes, then scale up. A failed 10-second render tells you less than a working 3-second one.

Wan 2.2 TI2V-5B files

The 14B models use a pair of diffusion files — a high-noise and a low-noise model, for example wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors and its low-noise partner — plus wan_2.1_vae.safetensors and the same UMT5 text encoder.

HunyuanVideo 1.5 files

An optional hunyuanvideo1.5_1080p_sr_distilled_fp16.safetensors upscales output to 1080p.

LTX-2.5 files

ComfyUI ships three native workflows for LTX-2.5: text-to-video, image-to-video, and first-last-frame (FLF2V), which interpolates between two keyframes. Note that the files ComfyUI uses are the int8_convrot builds, not the BF16 weights listed in Lightricks’ own repository:

Two things to know before you enable extras. The optional prompt enhancer, which expands a short prompt into a detailed one, needs a separate model of about 5GB and adds roughly 1–2 minutes per generation, so it is off by default. And the audio VAE is a separate file — it is what produces the synchronised soundtrack, so skipping it gets you a silent clip.

⚠️ The LTX-2.5 repository on Hugging Face is gated: you have to be logged in and accept the terms before the weights, or even the model card, will load.

Licences — check before you publish

• Wan 2.x: Apache 2.0. ComfyUI’s Wan documentation states the models “are based on the Apache 2.0 open source license and support commercial use”, and Wan’s own card adds that it claims no rights over what you generate. This is the least restrictive option here.
• LTX-2.5: LTX-2.x Community License. The licence is permissive with one revenue test: entities “with annual revenues of at least $10,000,000 (the ‘Commercial Entities’) are required to obtain a paid license for any use” other than a strictly non-commercial one. Below that threshold, commercial use is allowed. The licence is dated 11 August 2026 and covers every LTX-2.5 release since.
• HunyuanVideo 1.5: Tencent Hunyuan Community License. This one has a catch worth reading twice — the agreement states in its opening line that it “DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA”. Separately, any licensee above 100 million monthly active users must request a licence from Tencent. If you or your audience are in those territories, use Wan instead.

Making it fit

Video runs into VRAM limits far faster than images. In rough order of impact:

• Use the quantised or distilled build where the model offers one — FP8 for Wan, the int8 files for LTX-2.5, the distilled model for faster generation.
• Let ComfyUI offload. Its native offloading is what gets Wan 2.2 5B onto an 8GB card at all.
• Generate shorter clips at lower resolution, then upscale with the dedicated upscaler model rather than rendering big in one pass.

Our ComfyUI VRAM page explains the offload modes, and how to increase VRAM covers what actually helps on each kind of hardware. If you’re choosing a card for this, see our best GPU for AI guide. For still images in the same front end, see Stable Diffusion inpainting.

FAQ

Can I generate AI video locally on 8GB of VRAM?

Yes, with the smaller models. ComfyUI says Wan 2.2 TI2V-5B fits 8GB with its native offloading, and Wan states 8.19GB for Wan 2.1 T2V-1.3B. Expect slow generation and short clips.

Which local video model can I use commercially?

Wan 2.x, under Apache 2.0, is the simplest. LTX-2.5 allows commercial use below $10m in annual revenue. HunyuanVideo 1.5’s licence does not apply in the EU, the UK or South Korea at all.

Wan, HunyuanVideo or LTX?

Wan for the lowest VRAM and the most permissive licence. HunyuanVideo 1.5 if you have a 24GB card and the territory limits don’t affect you. LTX-2.5 if you want sound generated with the video.

How long does a local AI video take?

Minutes per clip on consumer hardware. Wan quotes about 4 minutes for a 5-second 480P clip with the 1.3B model on an RTX 4090, and under 9 minutes for 5 seconds of 720P with the 5B model. Offloading makes it slower.

Why are the video nodes missing after I load a workflow?

Usually because ComfyUI is behind. Its documentation notes that new core nodes often reach the nightly build before a stable release, so a workflow published alongside a new model may need nightly rather than the current stable version. The built-in templates make that concrete: each one declares a minimum ComfyUI version, and our guide to the 660 ComfyUI templates covers what ships and how much of it is cloud-only.

Sources

• ComfyUI docs: Wan 2.2, HunyuanVideo 1.5, LTX-2.5, Workflow templates
• Model cards: Wan2.2-TI2V-5B, Wan2.1-T2V-1.3B, HunyuanVideo-1.5
• Licences: Tencent Hunyuan Community License, LTX-2.x Community License
• Lightricks: LTX-2 repository (model files and variants)
• ComfyUI releases: Comfy-Org/ComfyUI

Updated 10 October 2026: ComfyUI release figures refreshed; model details as checked 17 September 2026. Memory, speed and licence details are as published by each vendor or by ComfyUI; BFWR does not run its own benchmarks.