Pick a model and a precision (FP8, GGUF…) to see how much VRAM it needs and whether your GPU can run it — or pick your GPU to see which models run. Estimates are for ComfyUI.
Estimates for ComfyUI (about ±20%). ControlNet, LoRAs and upscaling add more. Check each model's page before buying hardware.
It is the sum of the model weights (parameters × bytes per parameter for the chosen precision), the working memory during generation (how many latent tokens the image or video is split into × the model width), and a fixed part for VAE decoding. Model sizes and widths come from each model's official config.json and released weight files.
Image and video models first run a text encoder to read your prompt, then the diffusion model itself. When VRAM is short, ComfyUI swaps the two automatically. Only the first image is slower; generation speed itself is unchanged.
FP8 halves the file size with almost no visible quality loss. If VRAM is still short, step down through GGUF Q8 → Q5 → Q4. Q4 is usually fine, with slightly less fine detail.
These are estimates for ComfyUI and can be off by about ±20%. We checked them against the official numbers published by each model team and user reports. Video needs grow quickly with resolution and frame count, and ControlNet or LoRAs add more.
Running language models locally instead? Use the Local LLM VRAM Calculator.
Sources: each model's Hugging Face model card, config.json and released file sizes, and Black Forest Labs' FLUX.2 announcement (checked October 2026). Calibrated against official VRAM figures and user reports.