Do AI Artists Really Need an RTX 5090?

Most don't. A 16 GB card runs SDXL and FLUX.1 at 1024×1024 in ComfyUI, so many artists who make images for their own work can start there. The GeForce RTX 5090 earns its price when time matters or models get bigger. In one same-rig ComfyUI test it finished a FLUX.1-dev image in 8.84 seconds against 37.02 seconds on an RTX 5060 Ti 16GB, about 4.2 times faster. It also has the 32 GB needed for newer, larger models, video generation and training your own LoRAs without constant workarounds. The honest question is how many images you make in a day, and what your waiting time is worth.

How fast are different GPUs at AI image generation?

Source: computingforgeeks, ComfyUI v0.42.1 on rented cloud GPUs, 1024×1024, batch 1, warm mean of three runs with fixed seeds (measured 20 Oct 2026). The RTX 5070 Ti and 5080 were not measured in that test. Images per hour is our arithmetic (3,600 ÷ seconds per image) and ignores time spent loading models and adjusting prompts.

A second, older dataset tells the same story. In UL Procyon's SDXL FP16 test on a Threadripper rig, StorageReview measured 5.22 seconds per image on the 5090, 7.46 on the RTX 4090 and 8.81 on the RTX 5080 (January 2025). The 5090 even edged out the 96 GB RTX PRO 6000 (5.36 seconds) on the same test, because SDXL fits comfortably in 32 GB.

How much VRAM do current image models need?

• SDXL and SD 1.5: run on 8–12 GB cards, as the RTX 3060 result shows.
• FLUX.1-dev in fp8: peaked at 17.6 GB on the 5090 and 15.7 GB on the 5060 Ti 16GB, which managed only by running at its memory limit. Add LoRAs, ControlNets or higher resolutions and a 16 GB card starts offloading or failing.
• FLUX.1 in FP4 on RTX 50 cards: Black Forest Labs and NVIDIA say it needs only 10 GB and runs 2x faster on a 5090 than BF16 on a 4090. That's a vendor claim without published settings. Similar claims put FLUX.1 Kontext at 12 GB in FP8 and 7 GB in FP4, down from 24 GB.
• FLUX.2 (November 2025): a 32-billion-parameter model that NVIDIA says needs 90 GB to load completely and 64 GB even in ComfyUI's low-VRAM mode. FP8 cuts that by 40%, and weight streaming offloads the rest to system RAM. So FLUX.2 runs on a 5090 only with compromises, and on a 16 GB card it is a struggle.

If your images come out in seconds and your models fit, a faster GPU buys you time, not new abilities. Buy the 5090 when the models you want stop fitting.

Which cheaper GPUs make sense for AI art?

• RTX 5060 Ti 16GB: the entry point for FLUX.1 at 1024×1024 and comfortable for SDXL. It's slow on FLUX (about 37 seconds per image above), but usable for a hobbyist or someone making a few dozen images a day. Avoid the 8 GB version for AI work. Our RTX 5090 vs RTX 5060 Ti 16GB post covers it in depth.
• RTX 5070 Ti or 5080 (16 GB): faster than the 5060 Ti but no more memory, so the same model limits apply. We found no same-test image-generation figures for the 5070 Ti.
• A used RTX 3090 (24 GB): 23 seconds per FLUX image in the test above, with more memory than any 16 GB card. Used cards carry no warranty certainty and draw a lot of power.
• A used RTX 4090 (24 GB): sold for around US$3,500 on eBay in November 2026, which is no longer cheap.
• A GB10 system such as DGX Spark: 128 GB of unified memory fits large models, but it is much slower per image than a 5090. NVIDIA quotes 23 FLUX.1 Schnell images a minute at 1024×1024 and four steps with TensorRT FP4, a vendor figure for the fast four-step model. See RTX 5090 vs DGX Spark.

Street prices for GeForce cards have been unusually high and volatile through 2026. NVIDIA's US MSRPs are unchanged (US$1,999 for the 5090), but US street prices for a 5090 ran around US$6,500–8,500 in late October and November 2026. Date any price you compare.

When does the RTX 5090 earn its price for an artist?

1. Volume work. At about 407 FLUX images an hour against about 97, a studio iterating on client concepts saves hours a week.
2. Bigger and newer models. FLUX.2, larger editing models and multi-model ComfyUI graphs need the extra memory.
3. Training your own LoRAs and styles. Training needs much more memory than generating, and 32 GB gives room to train at higher resolutions without aggressive memory-saving settings.
4. Moving into video. Image-to-video models are where 16 GB runs out first. See can RTX 5090 run your own AI video studio?

Software matters as much as hardware. Blackwell cards need PyTorch 2.7 or later built for CUDA 12.8, and ComfyUI now calls for newer CUDA builds still. NVIDIA also says its NVFP4 checkpoints in ComfyUI are "up to" three times faster with 60% less VRAM on RTX 50 cards, a vendor claim worth testing on your own workflow. NVIDIA's Studio drivers are tuned for creative apps, but the company says either Studio or Game Ready drivers will run them. If your images feed 3D work too, see RTX 5090 for Blender.

Buying in Singapore

GeForce RTX 5090 and 5060 Ti cards are not in stock at SourceIT at the time of writing; ask us about availability and a quotation. In stock, the RTX PRO 4500 Blackwell offers 32 GB with ECC at S$8,880. The RTX PRO 4000 Blackwell has 24 GB at S$5,640. The NVIDIA DGX Spark 4TB is S$9,300, and the ASUS Ascent GX10 128GB / 1TB is S$10,350. For a compact all-in-one, the ASUS ROG NUC 16 with an RTX 5080 Laptop GPU (16 GB) is S$5,650.03. All prices listed 6 Nov 2026; prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Can a 16 GB GPU run FLUX? Yes, FLUX.1-dev in fp8 ran on an RTX 5060 Ti 16GB at 1024×1024, peaking at 15.7 GB. That leaves almost no room for LoRAs or ControlNets, and it took about 37 seconds per image in that test.

How much faster is an RTX 5090 than an RTX 5060 Ti for AI images? About 4.2 times on FLUX.1-dev fp8 (8.84 vs 37.02 seconds) and about 2.3 times on SDXL (4.84 vs 11.34 seconds) in one ComfyUI test, with the 5090 capped at 450 W.

Can an RTX 5090 run FLUX.2? With compromises. NVIDIA says FLUX.2 needs 90 GB to load completely; on a 5090 it relies on FP8 weights and ComfyUI's weight streaming to system RAM.

Is a used RTX 3090 a good budget option for AI art? It has 24 GB and ran FLUX.1-dev at about 23 seconds per image in one test. Weigh that against its age, power draw and the lack of a warranty you can rely on.

Should I buy an RTX 5090 Laptop GPU for AI art? It has 24 GB of VRAM, which helps, but it is a different and much slower chip than the desktop RTX 5090. Compare it with desktop options before choosing.