Wan 2.1 I2V 480P and Wan 2.1 I2V 720P are the same size on disk: 18.1 GB each at Q8_0. Yet on this site the 720P version needs about 2.5 GB more memory for the same file. That difference is the video itself.
Weights are only half the story
While sampling, the model holds a compressed version of every frame in memory — the latent — and works on all of them together, because each frame has to agree with its neighbours. An image model at 1024×1024 works on one picture. A Wan model at 832×480 for 81 frames works on about twenty compressed frames at once (Wan compresses time four-fold). The attention step, where every part of the clip looks at every other part, takes more and more time as the clip grows; with the memory-efficient attention ComfyUI uses, its memory grows roughly in step with the clip.
What moves the needle
- Resolution. 1280×720 has 2.3 times the pixels of 832×480. That is the jump between the 480P and 720P figures here.
- Frame count. 121 frames is 1.5 times 81. Longer clips cost memory roughly in step.
- The VAE decode at the end. Turning latents back into frames is a memory spike of its own.
VAE Decode (Tiled)fixes most out-of-memory errors that happen in the last seconds.
Ways to fit a big video model
- Use a smaller GGUF of the model — the same rule as for images. Which quant →
- Render at 480p and upscale afterwards.
- Render shorter clips and join them.
- Kijai's WanVideoWrapper nodes offer "block swap": some transformer blocks stay in system RAM and are swapped in as each step needs them. Slower, but it gets a 14B model onto small GPUs. ComfyUI's own Dynamic VRAM now does similar streaming by default on NVIDIA.
The working-memory figures on this site are estimates for each model's standard setting, listed on the methodology page. Measured values will replace them as benchmarks come in.