Video generation on your own GPU: VRAM, render time and what is realistic

An image model draws one picture. A video model draws a few dozen of them at once and has to keep them consistent, which makes it a different kind of problem. The memory goes up, the time goes up much faster, and a clip that looks short on a timeline can cost an hour on a modest card.

Why video is so much harder

A video model does not work on frames one by one. It compresses the whole clip into a single latent volume and denoises that as one long sequence, so every frame can attend to every other. That is what keeps motion coherent, and it is also the cost. A 5 second clip at 480p is a sequence of over 30,000 tokens, while a 1024 pixel image is about 4,000 on a model like Flux.

Attention compares every token with every other token, so the work grows with the square of the sequence. At image sizes it is a footnote to the model's own layers. On a clip it is most of the compute, and leaving it out made our first estimates several times too optimistic. This is the table that shows the shape, for Wan 2.1 1.3B on a card with 24 GB:

Output	Sequence	Memory	Time	Fits?
480p, 3 seconds	20,280 tokens	15 GB	5 min 14 s	yes
480p, 5 seconds	32,760 tokens	16 GB	12 min 40 s	yes
720p, 3 seconds	46,800 tokens	17 GB	24 min 53 s	yes
720p, 5 seconds	75,600 tokens	18 GB	1 h 2 min	yes
720p, 10 seconds	147,600 tokens	21 GB	3 h 52 min	yes

Going from 3 to 5 seconds multiplies the sequence by about 1.6 and the time by about 2.5. Going from 480p to 720p more than doubles the sequence, and the time grows nearly 5 times. These are the numbers to hold on to when you plan a project, since they mean that the first thing to cut is resolution, and the second is length.

Which models are practical

Three families cover most of what people run at home. Wan, in its 1.3B, 5B and 14B sizes, is the quality reference among open models. LTX-Video is built for speed and trades some quality for it. Stable Video Diffusion animates a still image into a short clip. The 1.3B Wan model is the one that makes video possible on a mid-range card, and its text encoder is bigger than its denoiser, which is why moving the encoder to the CPU saves so much:

Part of Wan 2.1 T2V 1.3B	Parameters	BF16	FP8	4-bit
transformer	1.42B	2.64 GB	1.32 GB	0.80 GB
text encoder	5.68B	10.6 GB	5.29 GB	3.20 GB
vae	0.13B	0.24 GB	0.24 GB	0.24 GB
All parts	7.227B	13.5 GB	6.85 GB	4.23 GB

Each one has its own native size and step count, and the table uses them. LTX-Video produces a clip in minutes on a mid-range card. The Wan models take a good deal longer for what is generally a better result.

Device	Memory	LTX-Video	Wan 2.1 T2V 1.3B	Wan 2.2 TI2V 5B
GeForce RTX 3060 12GB	12 GB	1 min 48 s GGUF Q8_0	35 min 14 s GGUF Q8_0	40 min 26 s GGUF Q4_K_M
GeForce RTX 4060 Ti 16GB	16 GB	1 min 4 s BF16	20 min 26 s GGUF Q8_0	23 min 28 s FP8
GeForce RTX 3090	24 GB	40 s BF16	12 min 40 s BF16	14 min 35 s GGUF Q8_0
GeForce RTX 4090	24 GB	19 s BF16	5 min 28 s BF16	6 min 20 s GGUF Q8_0

The 14B and larger models do not appear because they need more than one consumer card can hold, even at 8 bits, and run only with heavy offload. That is possible and measured in hours per clip. We would not call it practical.

What to do about it

Use the smallest model that gives you the quality you need, and do not start at 720p. Prototype at 480p and a few seconds, and render the final at a higher size only for the shot you keep. Move the text encoder to the CPU, since it runs once and removing it from the card frees gigabytes. Always use tiled decoding on video, because the decode step at the end is where clips that finished denoising run out of memory.

If you have 12 GB or less, plan on the 1.3B Wan model and LTX-Video, and treat the 5B as an overnight job. With 24 GB you can run the 5B comfortably and the 14B with offload, and it is the amount where video starts to feel like a tool and not an experiment. The 24 GB guide covers what else that card does well, and the video calculator lets you pick a model, a clip length and a card and see whether it fits.