Home › Image & video models › SDXL
New here: FLUX per-variant pages — schnell · dev · Kontext · FLUX vs SDXL
SDXL VRAM requirements — what is official, and what is not
Official statement: none. As of 2026-10-24 Stability AI publishes no VRAM requirement for SDXL — not on the model card,
not in Stability-AI/generative-models README (raw file retrieved 2026-10-24; the only “Low VRAM environment” passage in it belongs to the Stable Video Diffusion section and gives flags — --encoding_t=1, --decoding_t=1, lower --img_size — not a GB figure), not in SDXL paper (arXiv:2307.01952), full text retrieved 2026-10-24: “These changes result in a model size of 2.6B parameters in the UNet”; its architecture table lists “Text encoder: CLIP ViT-L & OpenCLIP ViT-bigG” and publishes no memory requirement. Anyone quoting “SDXL needs exactly N GB” is quoting a guess or one person’s rig.
The arithmetic we can honestly show
- UNet params, official: 2.6B (SDXL paper (arXiv:2307.01952), full text retrieved 2026-10-24: “These changes result in a model size of 2.6B parameters in the UNet”; its architecture table lists “Text encoder: CLIP ViT-L & OpenCLIP ViT-bigG” and publishes no memory requirement).
- Repo params, official metadata: huggingface.co/api/models/stabilityai/stable-diffusion-xl-base-1.0 → safetensors F32 params 2,567,463,684, retrieved 2026-10-24 — which is the paper’s 2.6B UNet, and matches it.
- fp16 weight floor: 2,567,463,684 × 2 bytes ≈ 4.78 GiB; fp32 ≈ 9.56 GiB — derived by us from those two official counts. A floor, not a running footprint.
- What stacks on top and is not officially quantified: the second text encoder pair (CLIP ViT-L + OpenCLIP ViT-bigG, named in the paper’s architecture table), the VAE decode pass, and activations at 1024². None of them has an official GB figure → no public data (we checked, 2026-10-24).
“Can my card run SDXL?” — answered as far as the sources allow
| Card | fp16 UNet weights alone (official params × 2 B) | Full SDXL footprint, officially stated |
|---|---|---|
| 4 GB | fp16 weights do not fit (4.78 GiB > 4 GB) | no public data (we checked, 2026-10-24) |
| 6 GB | fp16 weights fit, ~1.2 GiB left | no public data (we checked, 2026-10-24) |
| 8 GB | fp16 weights fit, ~3.2 GiB left | no public data (we checked, 2026-10-24) |
| 12 GB | fp16 weights fit, ~7.2 GiB left | no public data (we checked, 2026-10-24) |
| 24 GB | fp16 weights fit, ~19.2 GiB left | no public data (we checked, 2026-10-24) |
Middle column is arithmetic on an official count, so it is as true as that count. Right column is the honest gap: the number people actually want has never been published by Stability, so we leave it empty instead of inventing it. Testing without buying: RunPod RTX 5090 32GB @ $0.99/h is the cheapest ≥12 GB anchor in our cited table (modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending. Re-check 2026-10-22 (us): source table values unchanged since its stated 2026-09-29 check).
Measured-anchors (deliberately empty)
Run SDXL (or a refiner pass) at your resolution/precision, note the peak memory your driver reports, post it anywhere public — we cite reports with handle & date, and that beats any blog number including ours. No such citation is on this page yet.
FAQ
How much VRAM does SDXL need?
No official figure exists (we checked the Stability-AI/generative-models README and the SDXL paper on 2026-10-24 — neither states a memory requirement). Officially checkable: 2.6B UNet params → fp16 weights alone ≈ 4.78 GiB, fp32 ≈ 9.56 GiB. The two text encoders (CLIP ViT-L + OpenCLIP ViT-bigG) stack on top, and neither their parameter count nor their footprint is published → no public data (we checked, 2026-10-24).
Can SDXL run on a 4 GB card?
Not on unquantised fp16 weights, by arithmetic alone: the fp16 weight floor ≈ 4.78 GiB already exceeds a 4 GB card before encoders, VAE and activations. Whether some quantised or pruned community build runs on 4 GB is a measurement — we publish zero unverified numbers, ours or anyone else's.
What is SDXL's minimum VRAM?
Nobody official answers that question. The official README's only “Low VRAM environment” passage belongs to the Stable Video Diffusion section and hands you switches (--encoding_t=1, --decoding_t=1, lower --img_size), not a GB number. Treat any blog's “minimum 8 GB” as one person's rig, not a spec.
Does the SDXL refiner need more memory?
The paper introduces a refinement model used as “a post-hoc image-to-image technique” but publishes neither its parameter count nor its memory footprint → no public data (we checked, 2026-10-24). Base + refiner stacks two UNets; that much is arithmetic, the real peak still needs a measurement.
Is SDXL free for commercial use?
The repo metadata tags the licence as openrail++ (Hugging Face API, retrieved 2026-10-24). We did not verify the licence's commercial clauses today, so we say nothing about them — read the licence text yourself before shipping anything.
Siblings in the same honest series: FLUX.1 VRAM (no official number) · Wan 2.1 VRAM (official numbers do exist there) · LLMs: model pages · method & sources.