Video Can Be Generated Locally Too? Looking Into Wan 2.2, LTX-2 and FramePack
14 August 2026
On the machines here I generate text and images routinely. Video I had never touched.
I had simply assumed it was too heavy for a home PC. Then through August 2026 I kept running into talk of video generation running locally. Three names come up: Wan 2.2, LTX-2 and FramePack.
Curious, I looked into whether any of it would run on what I have.
This article is a summary of published information, not a record of running these myself. The numbers vary between sources, so I have kept them as ranges (checked 14 August 2026).
What local video generation actually does
The shape of it is much like image generation. You write an instruction in words, and a video comes out that follows it.
The difference is volume. A video is made of anything from a dozen to several dozen images per second, so producing the same duration takes tens of times the computation of a single image. On top of that, consecutive frames have to stay consistent with each other. That is why this was assumed to be out of reach at home.
All three names above have published weights, so you can download them and run them yourself. The thing people run them in appears to be ComfyUI — the node-based tool already familiar from image generation.
How the three differ
Here is what I found. Figures varied between sources, so I have left them as ranges.
| Name | Memory needed (VRAM) | Length it can make | Character |
|---|---|---|---|
| Wan 2.2 (14B) | 24GB or more. Around 16GB compressed | Up to 10 seconds, 16fps | The name that comes up most. Licensed so anyone can use it commercially (Apache 2.0) |
| Wan 2.2 (lighter 5B) | 8–12GB | Same | Within reach of a 12GB card |
| LTX-2 | Some sources say 32GB or more | 720p, 5 seconds | Sold on speed. Said to produce 5 seconds in under a minute |
| FramePack (7B) | around 6GB | up to 120 seconds | Light above all, and can make long clips |
Laid out like this, the three have clearly separate personalities.
Would any of it run here?
Mapping it onto what I own.
| Hardware here | Wan 2.2 (14B) | Wan 2.2 (5B) | LTX-2 | FramePack |
|---|---|---|---|---|
| RTX 3090 (24GB) | Just about reachable | Comfortable | Not enough | Comfortable |
| RTX 3060 (12GB) | Not enough | Looks reachable | Not enough | Comfortable |
| Intel Arc B580 (12GB) | Not enough | Capacity is there; support is the question | Not enough | Capacity is there; support is the question |
With a 24GB card, the larger Wan 2.2 looks like it comes into range. But some sources say “24GB or more" while others say “48GB minimum if uncompressed", so this clearly shifts with the settings. There is no knowing without loading it.
FramePack is the nearest of the three. If it really runs in about 6GB, a 12GB card has room to spare. That is where I would start.
As for the Intel Arc card, the capacity is there on paper, but whether the software supports it is a separate question. I have been caught by exactly that before, on the text-generation side. Best not to expect too much here.
[kimono_product id="16559″]
Things that snagged while reading
The numbers disagree between sources
For the same Wan 2.2 14B, I found “24GB or more", “48GB minimum if uncompressed" and “8–12GB if compressed to FP8" — three different answers.
Presumably it swings on how much the model is compressed and on the resolution and length being produced. Which means “you need N GB of VRAM" is the kind of statement that carries no meaning without its conditions attached. Numbers of this sort are not worth fully trusting until you have measured them under your own conditions.
The licensing is not uniform
For Wan 2.2 I saw both “anyone can use it commercially (Apache 2.0)" and descriptions of research-only restrictions. For the other two I did not find a clearly stated set of terms.
If you plan to put the output on a blog or anywhere public, this is something to confirm for yourself before running anything.
You need somewhere to put it all
Video models are large files, like text models. And the videos you produce pile up as well. Image-generation models alone already take 249GB here. Starting on video would add to that.
In summary: this looks usable, so I will measure it later
- Local video generation has settled around three names: Wan 2.2, LTX-2 and FramePack. All have published weights and appear to run under ComfyUI
- They pull in different directions. Wan 2.2 for quality, LTX-2 for speed, FramePack for lightness and length
- A 24GB card here looks like it brings the larger Wan 2.2 into range. Even a 12GB card reaches FramePack and the lighter Wan 2.2
- But the stated memory requirements disagree between sources. These are not numbers that mean anything without their conditions
- Licensing is not uniform either. Confirm it yourself before use
My assumption that this was too heavy for a home PC was, at least, wrong. It looks like it will run on what I have.
Which leaves only running it and measuring. Here is what I intend to measure.
| What to measure | Why |
|---|---|
| Whether each of the three runs on the 24GB and the 12GB card | Whether the published memory figures hold |
| Time to produce the same length at the same resolution | Whether the speed differences match what is claimed |
| Memory in use while generating | Where the overflow boundary actually sits |
| How the output looks | What gets traded away for speed |
I will start with the lightest, FramePack, and work upward once I know something runs. I will write it up when there are results.
Sources
Checked 14 August 2026. Both the software and the models get updated. Check memory requirements and licensing as they stand at the time you install.