Image Generation VRAM Calculator

How this is calculated

Weights: Component sizes are read from the actual .safetensors files in each model's Hugging Face repository, summed per component (transformer/UNet, text encoders, VAE). Parameter counts come from the repo's safetensors index.

Activations: Activation memory is an approximation, not a measurement: coefficient x megapixels x batch size. Diffusion activation use varies a lot with the attention implementation (SDPA, xformers, flash-attention), the sampler and the pipeline, so treat it as a planning figure with headroom, not a guarantee. The FLUX.2 models reuse the FLUX.1 coefficient (2.1 GB per megapixel): they share the rectified-flow transformer architecture, and Black Forest Labs publishes no activation figure of its own.

Peak VRAM is the resident weights for your chosen offload strategy, plus activations, plus about 0.6GB of framework and driver overhead. BFWR does not run a hardware test lab, so nothing here is a benchmark — the weight figures are exact and the activation figures are planning estimates.

Model files checked on 2026-09-12.

Common questions

How much VRAM do I need to run Flux locally?

It depends almost entirely on precision. At 1024x1024, batch 1: BF16 with everything resident is about 36.5GB. FP8 with the text encoder on CPU is about 14.9GB, which just fits a 16GB card. NF4 with the encoder offloaded is about 10.1GB, so a 12GB card manages it. An 8GB card is not enough for Flux at any setting in this calculator — SDXL is the better target there.

Why does Flux need so much more memory than SDXL?

Two reasons. The transformer is 11.9B parameters against SDXL's 2.57B UNet, and Flux loads T5-XXL as a text encoder, which is another 9.77GB at FP16. SDXL's two text encoders together come to about 1.64GB. At 1024x1024 SDXL peaks near 9.2GB at FP16 with everything resident, against 36.5GB for Flux at BF16.

Can I use Flux commercially?

Only FLUX.1 [schnell], which is Apache 2.0. FLUX.1 [dev] ships under a non-commercial licence, so it cannot be used for commercial output without a separate agreement with Black Forest Labs. Qwen-Image is also Apache 2.0 and is the largest open image model available.

Does resolution or batch size cost more memory?

Both, and they multiply. Activation memory scales with total pixels multiplied by batch size, so four images at 1024x1024 costs about the same as one image at 2048x2048. The weights stay constant, which is why large models are relatively cheaper to batch than small ones.

What is sequential CPU offload?

A mode where components are streamed into VRAM one at a time instead of all being resident, so peak memory is set by the largest single component rather than the total. Note that for Flux it buys very little over simply offloading the text encoder — about 0.2GB — because the 11.9B transformer is itself the largest component. Sequential offload helps most on models where the text encoder rivals the transformer in size, such as SD 3.5 Medium.

Are these numbers measured?

Partly. The weight sizes are exact, read from the real safetensors files in each model's Hugging Face repository. Activation memory is an approximation, because it varies with the attention implementation, sampler and pipeline. Plan with headroom rather than to the last gigabyte.

Per-model detail pages

Each model has its own page with the full component breakdown, every precision and offload combination, and the list of machines that can run it.

Most people run these models through ComfyUI, and its offload settings are what decide whether a given card copes — the ComfyUI VRAM page works the same figures through its three offload modes.

Which GPUs have the VRAM these models need, and what that costs right now.

• How to Run Stable Diffusion Locally: Hardware, VRAM and SetupCurrent models from SD 1.5 to Qwen-Image, and the VRAM each needs.Read the guide
• How to Increase VRAM: What Actually Works on PC, Laptop and MacIt depends where the memory physically lives.Read the guide
• Best Budget GPUs for Local LLMs16GB cards, used RTX 3090s and current street prices.Read the guide