How much VRAM do I need for FLUX and Stable Diffusion XL?

Short answer: Stable Diffusion XL’s FP16 weights are about 7 GB (CloudGPU model page, read 2026-10-23). FLUX.1 (dev or schnell) has about 24 GB of weights at 16-bit and about 12-16 GB at FP8 (CloudGPU model page, read 2026-10-23); with the text encoders and VAE loaded, the 16-bit pipeline is about 31 GB (Hugging Face Diffusers blog, read 2026-10-23). VRAM need depends on model, precision and workflow: larger resolutions, ControlNet, LoRAs, upscalers and larger batches add to it.

Published figures

From CloudGPU’s model pages (VRAM figures dated 2026-09-08; page re-read 2026-10-23). The FLUX figures refer to the main model weights, not the whole pipeline:

Model FP16 Lower precision
Stable Diffusion XL about 7 GB n/a
FLUX.1 dev about 24 GB (weights) about 12-16 GB at FP8
FLUX.1 schnell about 24 GB (weights) about 12-16 GB at FP8

What changes the number

  • Resolution and batch size: larger images and batches use more VRAM.
  • ControlNet, IP-Adapter, LoRAs: each adds memory.
  • Text encoders: FLUX uses large text encoders (the T5-XXL FP16 file alone is about 9.8 GB on Hugging Face, read 2026-10-23); they are loaded alongside the main model unless you offload them. The FLUX.1-dev files add up to about 34 GB (23.8 GB main model, 9.8 GB T5, 0.25 GB CLIP, 0.34 GB VAE; Hugging Face file sizes, read 2026-10-23), and the Hugging Face Diffusers blog (read 2026-10-23) reports about 31.4 GB loaded in BF16.
  • Precision: FP8 roughly halves the main model’s memory with some quality trade-offs.

Choosing a card

VRAM need depends on model, precision and workflow. Start from the figures above, add what your workflow loads, and check peak VRAM with nvidia-smi on a first run. Offloading the text encoders lowers VRAM use at the cost of speed.

Check https://cloudgpu.app/deploy for the cards you can start now.

Using the ComfyUI template

Check the template description on https://cloudgpu.app/deploy to see whether FLUX weights are included; if they are not, download the FLUX model, text encoders and VAE after the instance is ready. Instance storage is limited, so check the size first: the full 16-bit FLUX.1-dev files take about 34 GB together (Hugging Face file sizes, read 2026-10-23: 23.8 GB model, 9.8 GB T5 text encoder, 0.25 GB CLIP, 0.34 GB VAE), so FP8 files take less space if you also want other checkpoints. See how to choose a rental GPU for ComfyUI.

Billing reminder

Startup takes about 2 minutes. Billing is per minute, rounded up, 1-minute minimum, and the instance is billed while a model downloads. See how per-minute billing works.

Frequently asked questions

How much VRAM does SDXL need?

About 7 GB for the FP16 weights (the Hugging Face SDXL base file is 6.94 GB, read 2026-10-23). Larger resolutions, batches and add-ons need more.

How much VRAM does FLUX.1 dev need?

About 24 GB for the 16-bit weights (about 31 GB with text encoders and VAE loaded) and about 12-16 GB at FP8, per CloudGPU’s model page and the Hugging Face Diffusers blog (read 2026-10-23).

Is FLUX.1 schnell lighter than dev?

Not by much. CloudGPU’s published VRAM figures (read 2026-10-23) are the same, about 24 GB at 16-bit and about 12-16 GB at FP8, and the Hugging Face weight files are of similar size.

Will instance storage hold FLUX?

Check your instance’s storage size first. Full 16-bit FLUX.1-dev files total about 34 GB (Hugging Face file sizes, read 2026-10-23), which leaves limited space for other models; FP8 files are smaller.

Is FLUX included in the ComfyUI template?

Check the template description on https://cloudgpu.app/deploy. If the weights are not included, download them after the instance starts.

Sources (read 2026-10-23, Asia/Shanghai)

The Deploy page lists the cards you can start. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-10-23 (Asia/Shanghai).