Local image generation is forgiving on hardware. Local video is not. The moment you move from still frames to motion, VRAM becomes the wall everything runs into, because the model has to hold many frames in memory at once. Wan 2.2 and LTX-2 are the two open models most people reach for, and knowing what each actually demands saves you from buying the wrong card or chasing a resolution your GPU can never reach.

Quick Answer

Wan 2.2's smaller 5B variant fits comfortably on a 24GB card for 720p and runs on roughly 8 to 12GB when quantised, while the heavier 14B needs careful optimisation to fit consumer cards. LTX-2 is more demanding still, with an official minimum around 32GB for 720p full quality and roughly 12 to 24GB at 720p with FP8 quantisation. A 24GB card like the RTX 4090 is the sensible consumer target.

Wan 2.2: The More Forgiving Option

Wan 2.2 comes in several sizes, and the gap between them is large.

  • The 5B variant runs well on a 24GB card for 720p and fits roughly 8 to 12GB once quantised, making it the realistic choice for most consumer GPUs.
  • The smallest 1.3B variant squeezes onto about 4 to 6GB with a GGUF build, useful on very modest cards.
  • The 14B variant is heavier and spans a wide range depending on optimisation. With FP8 and the text encoder offloaded to system RAM, people fit it on cards in the mid-teens of gigabytes at 720p, but it is a tuning exercise, not a drop-in.

The single most effective trick for the 14B on consumer hardware is offloading the T5 text encoder to CPU RAM, which frees up a large chunk of VRAM for most of the generation. On an RTX 4090, expect roughly four to five minutes per short clip with Wan 2.2.

LTX-2: Faster Per Clip, Hungrier for VRAM

LTX-2 is the heavier lift on memory. Its official minimum sits around 32GB for full-quality 720p, though with FP8 quantisation it becomes workable in the 12 to 24GB range at 720p. Pushing to higher resolutions or full quality pulls it firmly into 24GB-plus and beyond.

The trade-off is speed: where it fits, LTX-2 tends to render a short clip noticeably faster than Wan, in the region of a minute and a half on a 4090 for a five-second clip. So the choice is partly memory budget, partly how much you value generation speed once the model fits. For comparing what current cards offer at each VRAM tier, the GPU best sellers are a fast reference.

Picking a Card for Local Video

VRAM capacity is the gate, full stop. If a model does not fit, no amount of raw speed helps. As a practical guide:

  • A 24GB card such as the RTX 4090 is the comfortable consumer sweet spot, running Wan 2.2 well and LTX-2 in quantised form at 720p.
  • 32GB opens up LTX-2 closer to its full-quality settings and gives Wan 2.2 14B more breathing room.
  • Below 16GB you are confined to the smaller Wan variants and aggressive quantisation, with longer generation times and lower ceilings.

Because these models keep evolving and the memory demands only grow, buying VRAM headroom is the decision that ages best. Builders assembling a dedicated machine for this often start from a purpose-specced AI PC rather than retrofitting a gaming rig.

Frequently Asked Questions

How much VRAM do I need for local AI video?

For a workable consumer experience, target 24GB. That runs Wan 2.2 comfortably and LTX-2 in quantised 720p form. Smaller Wan variants run on 8 to 12GB, but you trade resolution, quality, and speed as you go lower.

Can I run Wan 2.2 on an 8GB GPU?

Yes, but only the smaller variants. The 5B model fits roughly 8 to 12GB when quantised, and the tiny 1.3B build fits even less. The full 14B model is not realistic on 8GB without severe compromise.

Why does LTX-2 need so much more VRAM than Wan?

LTX-2's full-quality pipeline is simply heavier on memory, with an official minimum near 32GB for 720p. FP8 quantisation brings it down to the 12 to 24GB range, but it remains more demanding than Wan's smaller variants.

What is the T5 offload trick for Wan?

It moves the large T5 text encoder from the GPU into system RAM for most of the generation, freeing a substantial amount of VRAM. This is the single most effective way to fit Wan 2.2 14B onto consumer cards that otherwise could not hold it.

Should I prioritise VRAM or speed for video generation?

VRAM first, every time. If the model does not fit in memory, it either fails or crawls with heavy offloading. Once you have enough VRAM for your chosen model and resolution, faster compute then improves how quickly each clip renders.

Match the card to the model before you generate a single frame. Compare VRAM across current options in the graphics cards at Evetech and give your local video workflow the memory it needs.