Image & Video Model VRAM Calculator
Estimate peak memory for image and video diffusion models by resolution, frame count and weight precision. FLUX.1 dev needs about 12.1 GB in FP8 with model swapping, 7.3 GB with GGUF Q4, and 35.4 GB with everything loaded in FP16.
12.1 GB peak VRAM
FLUX.1 [dev] at 1024×1024 with FP8 weights, loading one model at a time (ComfyUI default on smaller cards). Keeping everything loaded needs 19.7 GB.
- Diffusion model 11.1 GB
- Text encoders 4.5 GB
- VAE 0.2 GB
- Peak activations 3.1 GB
- Overhead 0.8 GB
GPU fit
- RTX 4060 8GB62% on GPU
- RTX 3060 12GB93% on GPU
- RTX 4060 Ti 16GBFitswith swapping
- RTX 4090 24GBFitsall loaded
- RTX 5090 32GBFitsall loaded
- RTX 6000 Ada 48GBFitsall loaded
- RTX PRO 6000 Blackwell 96GBFitsall loaded
- Apple M4 Max 128GBFitsall loaded
The VAE decode is the peak here. Turn on tiled VAE decoding to cut it to about 0.8 GB.
Two numbers: swapping vs everything loaded
Image pipelines have three parts: text encoders, the diffusion model and the VAE. ComfyUI and Forge can load them one at a time, so peak memory is the largest single stage (with swapping). Keeping all three loaded is faster between generations but needs their sum (all loaded).
What changes the number most
- Weight precision: FP8 halves the diffusion model; GGUF Q4 cuts it to about 30% of FP16.
- Text encoder: T5-XXL (Flux, SD 3.5) is about 9 GB in FP16; loading it in FP8 or swapping it out matters on 12–16 GB cards.
- Resolution and frames: activation memory scales with the number of latent tokens. Video multiplies that by the number of latent frames.
- VAE decode: decoding large images or many frames can be the real peak. Tiled VAE decoding fixes most "out of memory at the last step" errors.
Activation figures are calibrated estimates. Treat results as ±20% and leave headroom. For LLMs use the LLM VRAM calculator.
Frequently asked questions
How much VRAM does Flux need?
FLUX.1 dev needs about 35.4 GB with everything loaded in FP16, about 12.1 GB in FP8 with model swapping, and about 7.3 GB with GGUF Q4 weights. 12 GB cards run it well with GGUF Q4–Q5.
Why does my image model run out of memory at the end?
The final VAE decode converts latents into full-resolution pixels and can need several GB at high resolutions. Enable tiled VAE decoding in ComfyUI or Forge.
Does batch size multiply VRAM?
It multiplies activation memory, not weights. Weights are shared, so batch 4 costs much less than four times the memory.