Stable Diffusion XL

Stable Diffusion3.5B paramsCreativeML Open RAIL++-MReleased 2023-07-26

The 3.5B-parameter base model that is still the most widely-supported open image model on Earth — every UI, every ControlNet, every LoRA trainer targets SDXL first, and it runs on far less VRAM than FLUX.

Last verified 2026-07 · Page created; VRAM tiers verified against Stability AI's own SDXL guidance and widely-reported community minimums.

What it is good at

SDXL 1.0, released July 2023, is a 3.5B-parameter UNet-based diffusion model — architecturally older and smaller than FLUX, but its head start means it has the deepest ecosystem of any open image model: ComfyUI, Automatic1111, and Forge all treat it as a baseline, and the LoRA/ControlNet/fine-tune community around it dwarfs anything built for newer models. Its two-encoder CLIP text conditioning (77-token limit per encoder) is less prompt-literate than FLUX's T5 encoder, but the tooling maturity and speed more than compensate for most use cases.

License is CreativeML Open RAIL++-M — open for commercial and personal use, with a short list of prohibited-use restrictions baked into the license itself (rather than a blanket non-commercial clause like FLUX.1 [dev]).

The hardware it wants

SDXL is dramatically lighter than FLUX: 8 GB of VRAM is comfortable for full-precision (FP16) generation at its native 1024x1024 resolution, and 6 GB is workable with attention-slicing/optimization tricks most UIs enable by default. 12-16 GB gives faster generation and headroom for larger batches or an SDXL Refiner pass, but is a comfort upgrade, not a requirement — this is the image model to pick when GPU budget is the binding constraint.

VRAM by quantization

Image models don’t quantize the same way LLMs do — practical VRAM guidance instead:

Minimum 6 GBComfortable 12 GB

Runs on

RTX 5060 (8 GB)Tight — reduced context
RTX 3070 (8 GB)Tight — reduced context
RTX 3060 Ti (8 GB)Tight — reduced context
RTX 4060 Ti (8 GB)Tight — reduced context
RTX 5050 (8 GB)Tight — reduced context
RTX 4060 (8 GB)Tight — reduced context
RTX 4070 Laptop (8 GB)Tight — reduced context
RX 7600 (8 GB)Tight — reduced context
RTX 4060 Laptop (8 GB)Tight — reduced context
Arc A750 (8 GB)Tight — reduced context
RTX 3080 (10 GB)Tight — reduced context
Arc B570 (10 GB)Tight — reduced context
RTX 5070 (12 GB)Fits comfortably
RTX 4070 Ti (12 GB)Fits comfortably
RTX 4070 Super (12 GB)Fits comfortably
RTX 4070 (12 GB)Fits comfortably
RX 7700 XT (12 GB)Fits comfortably
RTX 3060 (12 GB)Fits comfortably
Arc B580 (12 GB)Fits comfortably
RTX 5080 (16 GB)Fits comfortably
RTX 5070 Ti (16 GB)Fits comfortably
RTX 4080 Super (16 GB)Fits comfortably
RTX 4080 (16 GB)Fits comfortably
RX 9070 XT (16 GB)Fits comfortably
RTX 4070 Ti Super (16 GB)Fits comfortably
RX 9070 (16 GB)Fits comfortably
RTX 5060 Ti (16 GB)Fits comfortably
RTX 4090 Laptop (16 GB)Fits comfortably
RX 7900 GRE (16 GB)Fits comfortably
RX 7800 XT (16 GB)Fits comfortably
RTX 4060 Ti 16GB (16 GB)Fits comfortably
RTX A4000 (16 GB)Fits comfortably
Arc A770 (16 GB)Fits comfortably
RX 7900 XT (20 GB)Fits comfortably
RTX 4090 (24 GB)Fits comfortably
RX 7900 XTX (24 GB)Fits comfortably
RTX 5090 (32 GB)Fits comfortably
RTX 6000 Ada (48 GB)Fits comfortably

Unified-memory machines — Apple Silicon Mac Studio/MacBook Pro (Max/Ultra chips) and AMD Strix Halo mini-PCs — share system RAM and GPU memory in one pool, so a 64–128 GB configuration can hold models that would otherwise need multiple discrete GPUs. Throughput is bandwidth-bound like any GPU; it isn’t modeled in the table above yet.

Recommended builds

Three ways to size a machine for Stable Diffusion XL, from the smallest GPU that fits to the no-compromise option with headroom for the next model up.

BudgetSmallest GPU that fits at Q4 — runs it, no headroom for bigger contexts.
BalancedOne tier up — comfortable context length, room to run alongside other apps.
No-compromiseFits at Q8 or FP16, or leaves room to step up to the next model class.

Build it, or rent the compute?

Owning the hardware is a one-time cost that keeps paying off the more you actually use it — every run after the first is free, the data never leaves your machine, and the GPU still does everything else a PC does (gaming, editing, the rest of your workload) between AI sessions. Renting cloud GPU time trades that upfront cost for pay-as-you-go pricing with no idle hardware to maintain, which wins for occasional, bursty, or much-larger-than-consumer-VRAM jobs. For a model that fits comfortably in the VRAM tiers above and gets used regularly, owning the GPU tends to be the cheaper path over time — for anything larger than a single card can hold, renting is often the more practical starting point.

Frequently asked questions

Can I run SDXL on a 6 GB GPU?

Yes, workably — with the attention-slicing and half-precision optimizations most local UIs (ComfyUI, Automatic1111) apply automatically. 8 GB is the more comfortable floor for consistent full-resolution generation.

Should I use SDXL or FLUX.1 dev?

SDXL if your GPU is 8-12 GB, you want the widest LoRA/ControlNet ecosystem, or you need faster iteration. FLUX.1 dev if you have 12 GB+ (quantized) or 24 GB (full quality) and want stronger prompt adherence and photorealism — they are not strictly better/worse, just different tradeoffs.

Do I need the SDXL Refiner model too?

No — it is optional. The base model alone produces full-quality 1024x1024 images; the Refiner adds a second denoising pass some workflows use for extra detail, at the cost of additional VRAM and generation time.

Related models:Llama 3.1 8B