VRAM calculator for image and video models

Each model has a working-memory allowance for one reference setting (shown under the inputs): about one megapixel for most image models, a short clip for video models. That is the figure used everywhere else on this site, so at the reference setting the calculator gives exactly the same answer as the model and GPU pages.

When you change the setting, 30 % of that allowance is treated as fixed and the other 70 % is scaled in proportion to pixels × frames × batch. That matches how memory-efficient attention behaves (linear in the number of tokens), but real use also depends on the attention kernel, tiled VAE decoding and the ComfyUI version, so treat the result as a guide and leave some headroom.

LoRAs are counted at their file size. For a ControlNet or any other extra model, type its file size. Model weights, the 0.8 GB reserve and the fit rule are the same as on every page: how the numbers work.