Video needs so much more memory than images because a single 1024x1024 image is about one million pixels, while a 5-second clip at 24 frames per second is 120 frames: the model works on more than a hundred times as much at once, and has to keep every frame consistent with its neighbours, so it cannot make them one at a time. Three things drive memory and they multiply: model size and precision (full, FP8 or GGUF), resolution (720p has about 2.25 times the pixels of 480p) and clip length (roughly proportional). A model that fits at 480p for 3 seconds can run

out of memory

at 720p for 5 seconds on the same card.

This guide covers the open-weight video models you can download and run yourself. The figures are practical ranges from official model cards and community tests, checked on 6 October 2026; results vary with the

ComfyUI

version, the workflow and how much system RAM is available for offloading. For image models, see the companion guide on

VRAM requirements

by model; both are updated monthly.

Reference

Name	Type	What it is
Wan 2.1 T2V 1.3B	8 GB: 480p, ~5 s clips \| 12 GB+: comfortable	The 8 GB option. About 8.19 GB in use, 480p output, around 4 minutes per 5-second clip on an RTX 4090. Quality a clear step below the bigger models. Apache 2.0.
Wan 2.2 TI2V 5B	8 GB: GGUF Q8 with offloading, slow \| 12-16 GB: quantised at 720p \| 24 GB: full precision (official minimum)	Text or image to 720p at 24 fps. Alibaba lists 24 GB (an RTX 4090) as the minimum at full precision and under 9 minutes for a 5-second 720p clip. The Q8 GGUF is about 5.4 GB; ComfyUI offloads the rest to system RAM. Apache 2.0.
Wan 2.2 A14B (T2V and I2V)	12-16 GB: GGUF Q4-Q5 with offloading, short clips \| 24 GB: FP8 at 480p, 720p with the text encoder offloaded \| 32 GB: FP8 at 720p \| 80 GB: full precision	The best open quality. Mixture of experts: 27B parameters total, 14B active, one expert for the noisy early steps and one for detail; only one expert is in VRAM at a time. FP8 is about 14.3 GB per expert, GGUF Q4-Q5 about 10-11 GB. Official requirement 80 GB is for full precision with nothing offloaded. Apache 2.0.
HunyuanVideo 1.5	12-16 GB: ~14 GB minimum with offloading \| 24 GB: quantised at 720p \| 32 GB: comfortable \| 28-48 GB: full precision 720p	Tencent, November 2025, 8.3B parameters, 480p or 720p with a built-in upscaler to 1080p; 80 GB recommended for best quality. Not the original HunyuanVideo (13B), which needs 60-80 GB at full precision; many older guides mix the two up.
LTX-2 (19B) / LTX-2.3 (22B)	16 GB: draft previews in FP8 (distilled pipeline) \| 24 GB: FP4, clips up to ~97 frames \| 32 GB: FP8 (official minimum) \| 48 GB+: BF16 (~43 GB file)	Lightricks, weights January and March 2026. Video with synced audio, up to 4K, and fast: around 90 seconds for a 5-second clip on an RTX 4090. The trade is memory; Lightricks also recommends 100 GB+ of free disk.
System RAM	32 GB workable minimum \| 64 GB comfortable for 14B models on 12-24 GB cards	Offloading moves part of the model to regular RAM, so system memory matters more for video than for images.
Disk	fast NVMe SSD	Video models are tens of GB each; loading from a slow drive adds minutes to every model switch.
Wan 2.5-3.0, Kling, Veo, Runway, Seedance	API only	No downloadable weights, so no VRAM figure; you pay per second of video instead. Wan 2.2 is the newest Wan with weights (as of October 2026).

Video without a 24 GB card

BitVector Prism runs the Wan and LTX video models from this guide on its own GPUs: pick the model, upload a frame or type a prompt, get the clip in the browser. Nothing to install, no offloading.

Try BitVector

Step by step

Read your VRAM: Windows Task Manager > Performance > GPU > Dedicated GPU memory; nvidia-smi on Linux. Read your system RAM too; 32 GB is the floor for offloading.
Pick the model row for your card from the table: 8 GB means Wan 2.1 1.3B or Wan 2.2 5B GGUF; 12-16 GB adds Wan 2.2 14B GGUF and HunyuanVideo 1.5 with offloading; 24 GB is the common video card (5B full, 14B FP8, LTX-2 FP4); 32 GB runs everything at 720p in FP8.
Download the precision the model page on this site lists for your tier (FP8 for 24-32 GB, GGUF Q4-Q5 for 12-16 GB) rather than the BF16 original.
Start at 480p and short clips (49-81 frames). Get motion and composition right, then upscale or re-render at 720p. Going from 480p to 720p roughly doubles frame memory, so a workflow at 14 GB at 480p may need 24 GB or more at 720p.
For Wan 2.2 14B on 24 GB, move the T5/UMT5 text encoder to system RAM (CPU device on the loader, or ComfyUI's automatic offload). Without it a 720p clip can pass 24 GB partway through.
For anything over 5-6 seconds, generate in segments and chain them: the last frame of one clip becomes the start image of the next. Most open models are tuned for about 5 seconds and lose quality beyond it.
If a clip still does not fit, lower frames before resolution, use tiled VAE decode, and keep the browser and games off the GPU.
If you need Wan 2.2 14B or LTX-2 at 720p regularly and own less than 24 GB, run those on the cloud:
PirateDiffusion
runs the Wan,
LTX
and H3 models on server GPUs from Telegram;
BitVector
from a web app.

Examples

Wan 2.2 14B image-to-video on 24 GB (ComfyUI)

UNETLoader x2: wan2.2_i2v_high_noise_14B_fp8_scaled, wan2.2_i2v_low_noise_14B_fp8_scaled | CLIPLoader: umt5_xxl_fp8_e4m3fn_scaled, device: cpu | Load VAE: wan_2.1_vae

WanImageToVideo: 832x480, 81 frames | KSamplerAdvanced high (steps 0-10, cfg 3.5) -> KSamplerAdvanced low (steps 10-20, cfg 3.5) | shift 5 | VAE Decode (Tiled)

Wan 2.2 5B on an 8-12 GB card

UnetLoaderGGUF: wan2.2_ti2v_5B_Q8_0.gguf | CLIPLoader: umt5_xxl_fp8 (cpu) | 1280x704 is the native 720p shape; start at 832x480, 49 frames, 20 steps

python main.py --lowvram

The same clip on the cloud

/workflow /run:wan22-i2v slow dolly in, she looks up and smiles, soft wind in her hair /length:81 (reply to your image on PirateDiffusion)

Tips

Reported times for a single 5-second clip: Wan 2.1 1.3B 480p about 4 minutes on an RTX 4090; Wan 2.2 5B 720p under 9 minutes (official); Wan image-to-video about 12.7 minutes on a 4090 against about 7 on a 5090 (roughly 45 percent faster in one published test); LTX-2 about 90 seconds on a 4090. Offloading can make any of these several times slower.
The 80 GB figure Alibaba gives for Wan 2.2 A14B is for full precision with nothing offloaded; FP8 and GGUF plus the text encoder in system RAM is how everyone actually runs it.
Two 14B experts means two files: check the model page for which expert a
LoRA
was trained on (high-noise or low-noise); loading it on the wrong one does nothing.
LTX-2 trades memory for speed and audio; if you can fit FP8 on 32 GB it is the fastest open path to a finished clip with sound.
A 5090 launched at 1,999 dollars but sold for roughly 4,500-6,000 in October 2026 and still cannot hold the full-precision models. For occasional video, renting a 48 or 80 GB GPU by the hour, or a flat-rate cloud plan, is cheaper and much faster than offloading on a smaller card.
The Wan, Hunyuan and LTX family pages on this site list every downloadable variant with its file size; a variant needs its file size plus 3-4 GB free in VRAM to run at speed.

Wan 2.2 is the newest Wan you can run locally; the Wan family page tracks any change.

Wan 2.2 14B, LTX-2 and H3 from Telegram

PirateDiffusion hosts the large video models at full precision and sends the finished clip to your chat. Unlimited generation on a fixed monthly price, less than a month of 5090 street price per year.

Try PirateDiffusion

Questions

Can I make AI video with 8 GB of VRAM?

Yes, but only short, low-resolution clips. Wan 2.1 1.3B makes 480p clips in about 8 GB, and Wan 2.2 5B runs as a GGUF file with offloading. Wan 2.2 14B, HunyuanVideo and LTX-2 need more.

Is 24 GB enough for AI video?

For most people, yes. A 24 GB card runs Wan 2.2 5B at full precision, Wan 2.2 14B in FP8 (720p with the text encoder offloaded), HunyuanVideo 1.5 and LTX-2 in FP4. It is the most common card for local video work.

Why does Wan 2.2 say it needs 80 GB?

That is the figure for the 14B model at full precision with nothing offloaded. With FP8 or GGUF files and the text encoder in system RAM it runs on 24 GB, and on 12-16 GB more slowly.

Is an RTX 5090 worth it for AI video?

If you make video often, the extra 8 GB over a 4090 (32 against 24) is what gets you comfortable 720p with Wan 2.2 14B and FP8 LTX-2, and it is roughly 45 percent faster. At two to three times its 1,999 dollar launch price, renting is cheaper for anyone who does not generate video every day.

Does HunyuanVideo need more VRAM than Wan?

HunyuanVideo 1.5 (8.3B) needs less than Wan 2.2 14B: about 14 GB with offloading. The original HunyuanVideo (13B) needs much more, so check which version a guide is talking about.