RTX 5090 vs RTX 5080 for local Stable Diffusion in 2026

Local image generation has moved on considerably since the SDXL launch window. Flux.1 Dev is now the quality benchmark for anyone running models at home, SD3 Medium and Large sit alongside it, and SDXL remains the workhorse for high throughput jobs. The RTX 5090 sits at around £1,899 to £2,150 depending on partner board, and the RTX 5080 lands in the £999 to £1,150 range. The real question is whether the performance and VRAM headroom justify the two to one price gap for image work.

Raw specs that matter for diffusion

The 5090 ships with 32GB of GDDR7 on a 512 bit bus, delivering around 1,792 GB/s of bandwidth, 21,760 CUDA cores and 680 fifth generation Tensor cores boosting to 2.41 GHz. The 5080 has 16GB of GDDR7 on a 256 bit bus at roughly 960 GB/s, 10,752 CUDA cores and 336 Tensor cores boosting to 2.62 GHz. For diffusion inference the numbers that matter are VRAM capacity, memory bandwidth and Tensor core throughput at FP8 and FP4. The 5090 has almost double the bandwidth and exactly double the VRAM. The delivered gap in ComfyUI and Forge tends to be 1.7x to 1.9x depending on the workflow.

Flux.1 Dev performance in ComfyUI

Flux.1 Dev at full FP16 needs around 23GB to 24GB of VRAM with the T5 XXL text encoder loaded. On the 5090 you can hold the entire model in memory with room for a LoRA stack and generate a 1024 by 1024 image in roughly 8 to 10 seconds at 20 steps using Euler. On the 5080 the FP16 version will not fit without CPU offloading, which drags a single image to 22 to 28 seconds. The realistic path on 16GB is the FP8 quantised checkpoint at 12 to 14 seconds per image with minor quality loss. NF4 gets you to 9 to 10 seconds but the fidelity drop is visible on skin and text.

SDXL and SD3 throughput

SDXL is where the 5080 looks strongest because the model fits comfortably in 16GB with room for ControlNet and multiple LoRAs. A 1024 by 1024 image at 30 steps with DPM++ 2M Karras runs in 3.2 to 3.8 seconds on the 5080. The 5090 does the same job in 1.9 to 2.3 seconds. Batch of eight images pushes the 5080 to around 26 seconds while the 5090 completes it in 14 to 15 seconds because it can hold the whole batch in VRAM without swapping. SD3 Medium behaves similarly to SDXL on both cards. SD3 Large at FP16 is a different story: it wants 22GB plus, which rules the 5080 out of the full precision workflow entirely.

The batch size and resolution ceiling

This is the argument that ends most debates. On the 5090 you can render Flux.1 Dev at 1536 by 1536 native, then upscale with a 4x model in the same session without unloading anything. You can queue a batch of four Flux images or sixteen SDXL images and walk away. On the 5080 you are constantly making trade offs: unload the text encoder, drop batch size to 2, or accept quantisation. For a hobbyist that is tolerable. For a small studio or anyone iterating on client briefs, the 5080 becomes a bottleneck within a fortnight.

Browse Birmingham AV workstation and gaming rigs on eBay

Cost per image and power draw

The 5090 draws up to 575W under a diffusion load, the 5080 sits around 360W. At the UK domestic rate of roughly 27p per kWh in July 2026, amortising the card price over three years and assuming four hours of daily use, the 5090 works out at roughly 0.7p per SDXL image and 2.4p per Flux image. The 5080 comes in at 0.9p per SDXL and 3.1p per Flux. The gap is smaller than the sticker price suggests because the 5090 finishes each job faster.

Recommendation and the BAV build

If your workload is SDXL only and you generate under 300 images a week, the RTX 5080 is the sensible pick and leaves headroom for a better CPU or a second NVMe drive. If Flux.1 Dev or SD3 Large is in your pipeline, or you plan to move into video with Wan 2.2 or Hunyuan in the next twelve months, the 5090 is the correct card and the 16GB ceiling on the 5080 will frustrate you. The Birmingham AV high end creator rig pairs the RTX 5090 Founders Edition with a Ryzen 9 9950X3D, 64GB of DDR5 6400, a 2TB Samsung 990 Pro system drive and a 4TB WD Black SN850X for model storage, all in a Fractal North XL with a 1000W Corsair RM1000e. Built, tested and warrantied for twelve months. Configurations vary across our current listings, with 32GB and 96GB RAM options and case colours in black and walnut.

Frequently asked questions

Can I run Flux.1 Dev on a 16GB card at all

Yes, using the FP8 or NF4 quantised checkpoints with the T5 encoder offloaded to CPU. Expect 12 to 20 seconds per 1024 by 1024 image at 20 steps. Quality is close to FP16 on most subjects but text rendering and fine skin detail degrade noticeably.

Is the 5090 worth it if I only do SDXL

Only if throughput matters. The 5090 is roughly 1.7x faster on SDXL and lets you batch sixteen images at once instead of two. If you are a hobbyist generating hero images one at a time, the 5080 gives you 90% of the practical experience for 55% of the price.

Will the RTX 5080 handle video models like Wan 2.2

Barely, and only at reduced resolution. Video diffusion models sit at 20GB to 30GB VRAM for useful lengths. The 5080 forces heavy quantisation and short clip limits, while the 5090 handles 5 second 720p clips without compromise.

What about the RTX 5070 Ti as a budget option

The 5070 Ti at 16GB is a reasonable SDXL card at around £749 but its 672 GB/s bandwidth makes it noticeably slower on Flux than the 5080. If your budget forces the choice, save a little longer for the 5080 or step up to a used 4090 if stock allows.

About Birmingham AV

We have sold over 87,000 items on eBay since 2017 and hold 24,756 buyer feedbacks at 98.9% positive, which places us among the highest volume refurbished PC operations on eBay UK. Every system ships with a twelve month warranty, UK based support, and a full stress test report. Birmingham AV Ltd is registered at Companies House under number 12383651, VAT number GB 348755066, based in Bromsgrove, Worcestershire. Browse the current range at ebay.co.uk/str/midlandsav.