Five open video models, and what each one makes
An open video model is one whose weights you can download, so it runs on your own graphics card with no queue. A download can still ask for an account: the LTX-2.5 README says that on a 401 or 403 error you accept the model terms on Hugging Face and log in with a Read token. Once the files are on your disk, nothing runs on the maker's servers, so a shutdown like Sora's cannot switch them off, although each license still sets terms. The table lists five model families with the sizes and outputs each project's own README states.
One naming trap: LTX-2.5 replaced LTX-Video. The LTX-Video README says LTX-2.5, released on August 11, 2026, supersedes both LTX-Video and LTX-2, and that the old repository is no longer under active development. A new project starts from LTX-2.5.
| Model | Sizes the README lists | What one run makes |
|---|---|---|
| Wan2.2 | T2V-A14B and I2V-A14B (480P and 720P), TI2V-5B (720P) | TI2V-5B: text to video or image to video at 720P and 24 fps. The release also has S2V-14B (speech to video) and Animate-14B |
| Wan2.1 | T2V-1.3B (480P), T2V-14B | T2V-1.3B: a 5-second 480P clip in about 4 minutes on an RTX 4090, without quantization |
| LTX-2.5 | 22B transformer, full and distilled | Video and audio together. The quick start asks for 121 frames |
| HunyuanVideo-1.5 | 8.3B parameters. 480p and 720p models, then super resolution to 1080p | Text to video and image to video, 121 frames by default |
| CogVideoX | CogVideoX-2B, CogVideoX-5B, CogVideoX1.5-5B | 2B and 5B: 6 seconds at 8 fps, 720 by 480. 1.5-5B: 5 or 10 seconds at 16 fps |
How much video memory each model asks for
Video memory, or VRAM, is the memory on your graphics card, and it decides which model you can run. The local LLM hardware lesson explains VRAM and offloading, and both ideas carry over to video.
Read the chart as sizes, not as a speed test, because each project measures under its own settings. Tencent's 14 GB is with model offloading on. The CogVideo README gives its figures with every diffusers saving turned on, and says that without them peak memory is about three times higher. Wan2.2's 80 GB is for one A14B run on a single card with offloading, and its 5B model needs 24 GB.
LTX-2.5's README states no memory figure. For tight memory it suggests --quantization fp8-cast --offload {cpu, disk}. ComfyUI's README says its weight streaming runs even the biggest open models on as little as 4 GB of VRAM and 8 GB of RAM.
Sources: GitHub: Wan-Video/Wan2.2, README (read November 6, 2026), GitHub: Tencent-Hunyuan/HunyuanVideo-1.5, README (read November 6, 2026), GitHub: zai-org/CogVideo, README (read November 6, 2026), GitHub: Wan-Video/Wan2.1, README (read November 6, 2026)
Check the license before you use a clip for work
A model that is free to download is not always free for every use. The table gives each license in one line. Each project links the full text, which is worth reading when the clip is for a client or a business.
| Model | License | What it asks |
|---|---|---|
| Wan2.2 and Wan2.1 | Apache 2.0 | The README claims no rights over what you generate and says you are accountable for how you use the models |
| LTX-2.5 | LTX-2.x Community License for the model, Apache 2.0 for the code | A company whose yearly revenue passes the license's threshold needs a paid license for any use beyond non-commercial purposes |
| HunyuanVideo-1.5 | Tencent Hunyuan Community License | It does not apply in the European Union, the United Kingdom or South Korea. A company with over 100 million monthly active users must ask Tencent for a license |
| CogVideoX-2B | Apache 2.0 | The README puts the 2B model, with its transformers module and VAE, under Apache 2.0 |
| CogVideoX-5B | CogVideoX License | Free for academic research. Commercial use needs a registered basic commercial license, free up to 1 million visits a month |
Run Wan2.2 from its README, step by step
The Wan2.2 README makes a good first run because every step is a command you can copy. Its TI2V-5B model makes text-to-video and image-to-video at 720P, and the README says the command below runs on a single card such as an RTX 4090. The command and its prompt are the README's own.
python generate.py --task ti2v-5B --size 1280*704 --ckpt_dir ./Wan2.2-TI2V-5B --offload_model True --convert_model_dtype --t5_cpu --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage"- Check your card. The README says the TI2V-5B command needs at least 24 GB of VRAM, and the A14B commands at least 80 GB.
- Clone the project with
git clone https://github.com/Wan-Video/Wan2.2.git, then runcd Wan2.2. - Check that torch is 2.4.0 or newer, then run
pip install -r requirements.txt. Ifflash_attnfails to install, the README says to install the other packages first andflash_attnlast. - Install the downloader with
pip install "huggingface_hub[cli]", then fetch the model withhuggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir ./Wan2.2-TI2V-5B. The README shows this command for the T2V-A14B model, and the TI2V-5B repository name comes from its model table. - Run the command below. The options
--offload_model True,--convert_model_dtypeand--t5_cpulower GPU memory use. - On a card as large as the A14B models need, remove those three options to run faster, as the README says.
The README does not say where the file is saved. It says that, without specific optimization, TI2V-5B can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU. Its efficiency table is an image of total time and peak memory per GPU, so read it for your card.
Or use ComfyUI instead of a command line
ComfyUI is the app from the ComfyUI lesson: boxes joined on a canvas instead of commands. Its README lists video workflows for Wan 2.1 and 2.2, HunyuanVideo 1.5, CogVideoX and LTX-Video 2 and 2.3, and offers a desktop app for Windows, Linux and macOS. The Wan2.2 README dates its ComfyUI support to July 28, 2025, the day the ComfyUI account posted about it.
Lightricks points LTX users to its ComfyUI-LTXVideo repository.
Wan2.2 is now natively supported in ComfyUI on Day 0
- Install Comfy Desktop and make a first image, as the ComfyUI lesson shows. Choose Local at first launch.
- Open the video workflow library that the README links, find a ready-to-run workflow for the model you chose, and load it in ComfyUI. The ComfyUI lesson says a template knows which model files it needs.
- If ComfyUI reports missing model files, let it download the files it asks for, as the ComfyUI lesson shows for a template.
- Type your prompt, press Run, and keep the first clip short.
The post as X's own embed showed it on 5 November 2026, with the time in UTC. X cuts it short with Show more. Select the picture to open the post on X.
Prompts for a model on your own computer
Write the prompt as the prompt lesson teaches: one shot, one action, the camera and the light. Four projects add their own advice or a rewriting step on top.
| Model | What the project adds | Source |
|---|---|---|
| Wan2.2 | Prompt extension with --use_prompt_extend: a Qwen model rewrites the prompt, through Alibaba Cloud's Dashscope API (it needs a key) or a local Qwen model | Wan2.2 README |
| HunyuanVideo-1.5 | Rewriting is on by default (--rewrite) and needs a vLLM server. With no endpoint set, the pipeline runs without it | HunyuanVideo-1.5 README |
| LTX-2.5 | One flowing paragraph of up to 200 words. Its pipelines can enhance a prompt with the enhance_prompt parameter | LTX-2 README |
| CogVideoX | English only, up to 226 tokens (224 for 1.5-5B). The README says the model was trained on long prompts and points to a script that uses a large model to improve yours | CogVideo README |
Common mistakes
- Reading the memory chart as a speed ranking. The projects measure under different settings, and CogVideo says its figures assume every saving is on.
- Turning off the memory savings on a small card. The CogVideo README says peak memory is then about three times the table's value.
- Starting from an LTX-Video v0.9.x tutorial. LTX-2.5 supersedes that model, so look for instructions written for LTX-2.5.
- Skipping the license. HunyuanVideo-1.5's does not apply in the European Union, the United Kingdom or South Korea, and CogVideoX-5B needs registration for commercial use.
- Writing a CogVideoX prompt in another language. Its README says the model takes English only.
Try it yourself
- Look up how much video memory your graphics card has, then pick the largest model on the chart that fits.
- Open that model's license and write one line on what you may do with a clip.
- Run the README's prompt about the cats in boxing gloves, then change only the subject.
- Take a shot you wrote in the last lesson and run it in two models, or in two ComfyUI templates.
- Write down which memory-saving options you used, so the next run starts from them.
Check: the clip plays, it matches the shot you wrote, and you can say its license in one line.
Next course: AI voice and music
Models such as LTX-2.5 and Veo make sound with the picture, but most video still needs a voice or a song laid under it. AI voice and music covers speech, voice cloning, songs and dubbing, starting with text to speech.
Check yourself
Which Wan2.2 model does the README say runs on a GPU with at least 24 GB of VRAM, such as an RTX 4090?
Show answer
B. TI2V-5B The TI2V-5B command is the one the README says needs at least 24 GB. The A14B commands need at least 80 GB.
Whose license says it does not apply in the European Union, the United Kingdom or South Korea?
Show answer
A. HunyuanVideo-1.5 The Tencent Hunyuan Community License opens by saying so. Wan2.2 and CogVideoX-2B use Apache 2.0.
What does the LTX-Video README say about LTX-2.5?
Show answer
C. It supersedes LTX-Video and LTX-2 The README says LTX-2.5, released on August 11, 2026, supersedes both, and that the old repository is no longer under active development.
You turn off the diffusers memory savings for CogVideoX. What does its README say happens?
Show answer
D. Peak memory is about three times the table value, and speed rises The README says peak memory becomes about three times the table's value, while speed rises by about three to four times.
Frequently asked questions
What are the best open source AI video models?
This lesson covers Wan2.2, LTX-2.5, HunyuanVideo-1.5 and CogVideoX, all with downloadable weights and a README that shows how to run them. The best one for you depends on your video memory and the license you can accept.
What graphics card do I need for AI video?
Each project states its own need. Find the smallest figure its README gives, match it to your card's VRAM, and turn on the project's memory-saving options when you are close to the limit.
Can I fine-tune an open video model?
Some projects ship training tools. The LTX-2 repository has an ltx-trainer package for LoRA, full fine-tuning and IC-LoRA, and CogVideo has a finetune folder for LoRA. The LTX license counts fine-tuned weights as derivatives, so read it first.
Is there an open source alternative to Sora?
There are several open models, such as Wan2.2 and LTX-2.5, that you run yourself. They are different models from Sora, with their own limits, so compare their clips before you rely on one.
Do open video models make sound?
LTX-2.5 does: the LTX-Video README describes synchronized audio and video. Wan2.2 has a separate speech-to-video model, S2V-14B, which takes an audio file as input.
What is a distilled video model?
It is a version made to need fewer steps, so it runs faster. LTX-2.5's distilled transformer, which its quick start uses, runs in far fewer steps, and HunyuanVideo-1.5's step-distilled model makes 480p image-to-video in 8 or 12 steps.
Sources
- GitHub: Wan-Video/Wan2.2, README https://github.com/Wan-Video/Wan2.2
- GitHub: Wan-Video/Wan2.1, README https://github.com/Wan-Video/Wan2.1
- GitHub: Lightricks/LTX-2, README https://github.com/Lightricks/LTX-2
- GitHub: Lightricks/LTX-Video, README https://github.com/Lightricks/LTX-Video
- GitHub: Lightricks/LTX-2, LTX-2.x Community License https://github.com/Lightricks/LTX-2/blob/main/LICENSE-2_x
- GitHub: Tencent-Hunyuan/HunyuanVideo-1.5, README https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5
- GitHub: Tencent-Hunyuan/HunyuanVideo-1.5, Tencent Hunyuan Community License https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/LICENSE
- GitHub: zai-org/CogVideo, README https://github.com/zai-org/CogVideo
- Hugging Face: THUDM/CogVideoX-5b, CogVideoX License https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE
- GitHub: comfyanonymous/ComfyUI, README https://github.com/comfyanonymous/ComfyUI
- X: ComfyUI, July 28, 2025 https://x.com/ComfyUI/status/1949802127066628370