Best GPU for Stable Diffusion in 2026 (SD 1.5, SDXL, FLUX)

Stable Diffusion is the easiest large-model workload to host: SD 1.5 will run on almost anything with a discrete GPU, and SDXL is comfortable on a 16 GB card. Where it gets interesting is FLUX.1 (Black Forest Labs’ successor lineage) — that’s where VRAM and FP8 hardware start to matter. This guide picks the right card across our 12-SKU catalogue for every variant from SD 1.5 through FLUX.1 dev.

Top picks at a glance

Model VRAM requirements

Before picking a card, know what you’re running. The Stable Diffusion family spans an order of magnitude of VRAM appetite:

• SD 1.5 — ~4 GB at FP16. Runs on essentially any discrete GPU sold this decade. The 8 GB cards in our catalogue handle it without quantisation.
• SDXL 1.0 — ~8 GB at FP16, ~4 GB at FP8. An 8 GB card can run it, but with refiner loaded you’ll want 12 GB. 16 GB is comfortable.
• SDXL Turbo — same memory footprint as SDXL, but single-step inference makes it dramatically faster on lower-end cards. Same hardware advice applies.
• FLUX.1 schnell — ~12 GB at FP16, ~6 GB at FP8. Needs a 12 GB+ card at full precision; FP8 lets it fit on 8 GB with care.
• FLUX.1 dev — ~24 GB at FP16, ~12 GB at FP8. This is the hard cut-off. FP16 needs a 24 GB+ card; FP8 brings it down to a 16 GB card with enough headroom for a LoRA or two.

Add 2–4 GB on top of those numbers for ControlNet, IP-Adapter, multiple LoRAs, or higher-resolution latents. ComfyUI workflows with several active nodes can easily double the working set.

GPU-by-GPU recommendations

NVIDIA cards

The RTX 3090 24 GB at £159/mo is the budget VRAM king. It just fits FLUX.1 dev FP16, runs every SDXL workflow you’ll throw at it, and has enough memory bandwidth to stay competitive on raw throughput. The catch: Ampere has no FP8 hardware acceleration, so you cannot exploit the FP8 speed-ups newer cards see. For a hobbyist who wants every model to run, it’s still excellent value.

The RTX 5060 8 GB at £99/mo is the cheapest “image-gen-capable” NVIDIA card. SD 1.5 is comfortable, SDXL works at FP8, FLUX.1 schnell fits at FP8 with tight headroom. FLUX.1 dev is off the menu. Step up if you want it.

The RTX 3050 8 GB at £79/mo is genuinely a hobbyist box: SD 1.5, lower-resolution SDXL, basic ComfyUI workflows. If you only ever generate SD 1.5 outputs, it’s the cheapest entry into the catalogue.

The RTX 5060 Ti 16 GB at £119/mo is our recommended starter for serious Stable Diffusion work. It has Blackwell FP8 hardware, 16 GB of VRAM, and the price is right. SDXL is comfortable, FLUX.1 schnell fits at FP16, and FLUX.1 dev runs at FP8 with enough room for one LoRA. It’s the card we point most new ComfyUI users at.

The RTX 5080 16 GB at £189/mo is what you upgrade to when SDXL renders feel slow. Same VRAM as the 5060 Ti, but materially faster. FLUX.1 dev FP8 has more breathing room here.

The RTX 4090 24 GB at £289/mo runs everything in our list at full precision. Ada Lovelace FP8 is good (though not as fast as Blackwell), and 24 GB makes FLUX.1 dev FP16 actually pleasant rather than barely-fits. It’s a strong pick if you specifically need FLUX.1 dev FP16 but don’t want to pay 5090 prices.

The RTX 5090 32 GB at £399/mo is the catalogue’s image-generation flagship. Blackwell FP4 hardware, 32 GB of VRAM, and the highest tensor throughput of any GeForce-class card we host. FLUX.1 dev FP16 is comfortable with ControlNet and LoRA stacks loaded; SDXL renders are roughly half the wall-clock of the 4090. If your business depends on render throughput, this is the card.

The RTX 6000 PRO 96 GB at £899/mo is for batch jobs and parallel pipelines. 96 GB lets you keep SDXL, FLUX.1 dev and a stack of LoRAs hot in memory simultaneously, run batch sizes the consumer cards can’t touch, and serve multiple ComfyUI instances off one box. ECC memory matters if you’re running 24/7 production. Most single-user image-gen workloads don’t need it, but for a small studio it pays for itself.

AMD cards

The Radeon RX 9070 XT 16 GB at £129/mo is the entry-tier AMD pick. SDXL works fine on ROCm, FLUX.1 schnell runs at FP8, FLUX.1 dev FP8 is tight but possible. The caveat is ROCm maturity: the NVIDIA path has more turnkey ComfyUI nodes, more pre-built FLUX wheels, and fewer “compile-it-yourself” moments. If you’re an AMD shop already, it’s a credible option.

The Radeon AI Pro R9700 32 GB at £199/mo is the best-value AMD card for FLUX.1 dev. 32 GB unlocks FP16 inference and gives you batch headroom — you’re paying less than the 4090 and getting more VRAM. Same ROCm caveat applies: budget time for stack setup if you’re new to ROCm.

Intel and unified-memory options

The Arc Pro B70 24 GB at £179/mo runs SDXL well on oneAPI / IPEX. FLUX support is improving but lags NVIDIA — viable for hobbyists curious about the Intel stack, less ideal as your only image-gen box.

The Ryzen AI MAX+ 395 at £299/mo is the unusual entry: 128 GB of unified memory shared with the CPU. Not the fastest card per render, but uniquely capable of holding gigantic batches or several models in memory at once. Niche, but useful for batch generation pipelines where wall-clock per-image matters less than throughput per pound.

SDXL throughput estimates

Rough wall-clock estimates for a single 1024×1024 SDXL image, 30 steps, no refiner. Treat these as order-of-magnitude — your workflow, scheduler and node graph will move them around:

The interesting jump is 4060 to 5060 Ti — half the wall-clock at a small price premium, because Blackwell FP8 actually accelerates the workload. The 4090-to-5090 jump is similar: FP4 plus more bandwidth roughly halves render time again.

Decision tree

• If you only ever run SD 1.5, pick the RTX 3050 (£79) — every penny above that is wasted on this workload.
• If you want SDXL on a budget, pick the RTX 5060 (£99) for 8 GB FP8, or step up to the RTX 5060 Ti (£119) if you might add FLUX later. The £20 delta is the best-spent £20 in the catalogue.
• If you want FLUX.1 schnell comfortably, pick the RTX 5060 Ti (£119) — it’s the cheapest card with enough VRAM and FP8 hardware to make schnell pleasant.
• If you want FLUX.1 dev at FP8, pick the RTX 5080 (£189) — 16 GB plus Blackwell FP8 is the right shape for this workload.
• If you want FLUX.1 dev at FP16, pick the RTX 4090 (£289) for the cheapest FP16 path, or the RTX 5090 (£399) for materially better throughput and ControlNet headroom.
• If you’re running batch / studio workloads, pick the RTX 6000 PRO (£899) — 96 GB is the difference between “serve one user” and “serve a team”.
• If you’re already an AMD shop, pick the Radeon AI Pro R9700 (£199) — 32 GB on ROCm is the best AMD value for FLUX.1 dev.

Conclusion

Stable Diffusion is forgiving — the RTX 5060 Ti at £119/mo will cover 80% of users from SD 1.5 through FLUX.1 schnell. If you need FLUX.1 dev at full precision, jump to the RTX 5090 at £399/mo. For everything FLUX-specific, our deeper dive lives at best GPU for FLUX.