Open-weight image-to-video models compared: Wan 2.2, HunyuanVideo, LTX

Wan 2.2, HunyuanVideo and LTX-Video are three open-weight model families with image-to-video options. For a 24 GB card, Wan 2.2 TI2V-5B is the only variant with an official 24 GB claim. Licenses differ sharply, and several official figures conflict or remain unconfirmed.

We have not run these models ourselves. All figures come from the official repos, model cards and docs linked below.

At a glance

Each cell reflects the official source for that model. “Not confirmed” means the sources we used did not state it.

Wan 2.2

The Wan 2.2 repo lists a July 28, 2025, release under Apache 2.0. It describes T2V-A14B and I2V-A14B as mixture-of-experts models with 27B total and 14B active parameters, at 480P and 720P.

TI2V-5B is a dense model that takes text and an image. Its model card lists 720P output at 1280×704 or 704×1280 and 24 fps.

The card also says it makes “a 5-second 720P video in under 9 minutes” on one consumer GPU.

VRAM conflict. The repo says TI2V-5B needs “at least 24GB VRAM (e.g, RTX 4090 GPU).” The ComfyUI Wan 2.2 docs say it “should fit well on 8GB vram with the ComfyUI native offloading.”

For A14B, the repo states “at least 80GB VRAM” on a single GPU. ComfyUI documents files for all three variants, though its 14B page mixes fp16 and fp8 file names.

Our Wan 2.2 ComfyUI setup guide lists every file and folder.

HunyuanVideo

HunyuanVideo (text-to-video)

The HunyuanVideo repo lists weights on December 3, 2024, and FP8 weights on December 18, 2024. It is text-to-video only, at 540p and 720p, with 129 frames by default.

The repo gives a 60GB memory peak at 720p with 129 frames, and 45GB at 544p. It recommends an 80GB GPU.

HunyuanVideo-I2V

The HunyuanVideo-I2V repo lists a March 6, 2025, release, with fixed weights on March 7, 2025. It supports up to 720p, and describes 129 frames as 5 seconds.

The repo gives a 60GB memory peak at 720p.

License warning. The model uses the “Tencent Hunyuan Community License Agreement.” Its license text says it “does not apply in the European Union, United Kingdom and South Korea.”

In plain terms, the license grants no rights to users in those regions. Companies above 100 million monthly active users must also request a separate license. This is our reading, not legal advice.

HunyuanVideo-1.5

The HunyuanVideo-1.5 repo describes an 8.3B-parameter model. It supports image-to-video at 480p and 720p, with an upscaler to 1080p, and 121 frames by default.

It states a minimum of “14 GB (with model offloading enabled).” The release date is unclear, because the sources give both November 20 and November 21, 2025. We did not confirm its license for this guide.

HunyuanVideo in ComfyUI

The ComfyUI HunyuanVideo docs list these files for the original models.

• clip_l and llava_llama3_fp8_scaled go in text_encoders/.
• hunyuan_video_vae_bf16 goes in vae/.
• hunyuan_video_t2v_720p_bf16 goes in diffusion_models/ for text-to-video.
• For image-to-video, llava_llama3_vision goes in clip_vision/.
• Add hunyuan_video_image_to_video_720p_bf16 (v1) or hunyuan_video_v2_replace_image_to_video_720p_bf16 (v2) to diffusion_models/.

LTX-Video and LTX-2

LTX-Video versions and licenses

The LTX-Video repo lists v0.9.0, with text-to-video and image-to-video, on November 21, 2024. A 13B v0.9.7 followed in May 2025.

v0.9.8 arrived on July 16, 2025, in 13B and 2B distilled versions. Since July 2025, the repo says the model “supports up to 60 seconds of video.”

Licensing changed across versions. Per the model card, the code is Apache-2.0, v0.9.5 weights are OpenRail-M, and weights from 0.9.6 use the “LTX-Video-Open-Weights-License.”

LTX-Video resolution and memory

The repo gives a default of 1216×704 at 30 fps. It says results are best below 720×1280 and under 257 frames.

The 2B distilled model is described as “ideal for light VRAM usage,” with no number given. The repo also claims the 13B distilled LoRA128 “requires only 1GB of VRAM.”

Unverified claim. That 1GB wording is ambiguous, and we have not been able to confirm what it covers. Treat it as a claim, not a requirement.

The repo also links a community Q8 build running 720×480 with 121 frames on an RTX 4060 with 8GB. That is a community result, not an official figure.

The official system requirements page is written for LTX-2.5. It asks for an “NVIDIA GPU with a minimum 32GB+ VRAM,” 32GB of RAM and 100GB of disk.

LTX-2

Lightricks announced LTX-2 on October 23, 2025. Its model card describes 19B parameters and gives an example of 121 frames at 24 fps.

It uses the “ltx-2-community-license-agreement.” We could not confirm the date its weights were released. Lightricks describes LTX-2 as “built into ComfyUI core.”

Safety filters

No built-in content filter is documented on the official pages for any of these models. The LTX-2 card warns that the model “may generate content that is inappropriate or offensive.”

With local models, the rules are yours. Never use photos of real people, and read our consent policy and our guide to real-person likeness law.

Which to try first on a 24 GB card

This order follows the official memory figures only. It is not based on our own tests.

1. Wan 2.2 TI2V-5B. It is the only model here with an official 24 GB claim, and ComfyUI says it can run in less with offloading. It is also Apache 2.0, and ComfyUI publishes its files.
2. HunyuanVideo-1.5. Its repo states a 14 GB minimum with offloading, which sits below 24 GB. Check its license and release notes first, since we could not confirm either.
3. LTX-Video. Lighter variants exist, but the 2B model has no official VRAM number. The official 32GB+ figure is for LTX-2.5, so treat LTX on 24 GB as experimental.

Wan 2.2 A14B, HunyuanVideo and HunyuanVideo-I2V list official figures well above 24 GB. Per their repos, those are 80GB, and a 60GB peak at 720p.

To compare running costs, use the cost-per-clip calculator or read cloud vs local cost per clip.

Sources and dates

Facts checked against the linked official pages on October 31, 2026. Open-weight releases move fast, so check each repo before you download.