RTX 4090 vs RTX 4090 48G vs RTX 5090 for image generation

Short answer: Choose by VRAM first. The RTX 4090 has 24 GB, the RTX 5090 32 GB and the RTX 4090 48G 48 GB on the same chip as the RTX 4090 (NVIDIA product pages and CloudGPU /pricing, read 2026-11-03). SDXL and FP8 FLUX.1 fit on all three. The full 16-bit FLUX.1-dev pipeline is about 31.4 GB loaded in BF16 (Hugging Face Diffusers blog, read 2026-10-23), so it needs FP8 or offloading on 24 GB, is close to the limit on 32 GB, and has room on 48 GB. When the model fits on both, the RTX 5090 has more memory bandwidth (1,792 vs 1,008 GB/s) and FP16 compute.

Specs side by side

RTX 4090 RTX 4090 48G RTX 5090
Architecture Ada Lovelace Ada (same chip as RTX 4090) Blackwell
VRAM 24 GB GDDR6X 48 GB 32 GB GDDR7
Memory interface 384-bit n/a 512-bit
Memory bandwidth 1,008 GB/s 1,008 GB/s 1,792 GB/s
FP16 TFLOPS (CloudGPU /pricing) 165 165 209
CUDA cores 16,384 same chip as RTX 4090 21,760
Tensor Cores 4th generation 4th generation 5th generation (NVIDIA lists FP4 support)
Compute capability 8.9 8.9 12.0

Sources: NVIDIA’s RTX 4090 and RTX 5090 pages, NVIDIA’s CUDA GPUs page, and CloudGPU’s spec comparison taken from /pricing (all read 2026-11-03). The RTX 4090 48G is not listed on NVIDIA’s product pages; its figures are the ones CloudGPU publishes.

Which workloads fit

Figures are file sizes and published guidance, not CloudGPU measurements. VRAM need depends on model, precision and workflow.

Workload What it needs RTX 4090 (24 GB) RTX 5090 (32 GB) RTX 4090 48G
SDXL about 7 GB FP16 weights fits fits fits
FLUX.1-dev FP8 checkpoint 17.25 GB file; about 12-16 GB weights fits fits fits, more room for LoRAs and batches
FLUX.1-dev full 16-bit about 31.4 GB loaded in BF16 needs FP8 weight_dtype or offloading close to the limit; offload the text encoder if it does not fit fits with headroom
Wan 2.2 TI2V-5B ComfyUI: “should fit well on 8GB vram” with native offloading fits fits fits
Wan 2.2 14B (T2V / I2V) two 14.29 GB FP8 experts plus a 6.74 GB FP8 text encoder in ComfyUI; the Wan repo’s own script says “at least 80GB VRAM” relies on offloading relies on offloading more room than the other two

Sources for the table: CloudGPU model pages (SDXL, FLUX FP8), Hugging Face file listings for Comfy-Org/Flux-1-Dev and Comfy-Org/Wan-2.2, the ComfyUI Wan 2.2 tutorial and the Wan2.2 README, read 2026-11-03.

What the numbers mean for image work

  • VRAM decides whether it runs without offloading. When the whole pipeline does not fit, ComfyUI moves parts of it to system RAM and each step slows down. More VRAM also leaves room for LoRAs, ControlNet, upscalers and larger batches.
  • Bandwidth and compute decide speed once it fits. The RTX 5090 has about 78% more memory bandwidth than the RTX 4090 and a higher FP16 figure (209 vs 165 TFLOPS). The RTX 4090 48G has the same compute and bandwidth as the 24 GB RTX 4090: it buys capacity, not speed.
  • Precision support. Ada (RTX 4090 and 4090 48G) and Blackwell (RTX 5090) both run FP8 checkpoints. NVIDIA lists FP4 for the RTX 5090’s 5th-generation Tensor Cores; whether your workflow uses FP4 depends on the model files and nodes you install.
  • Software for the RTX 5090. Compute capability 12.0 needs a PyTorch build for CUDA 12.8 or newer; older builds fail with “no kernel image”. See how to fix CUDA “no kernel image” on the RTX 5090.

How to decide

  1. SDXL, FP8 FLUX, Wan 2.2 5B: all three cards run them. Compare the hourly rates on /pricing with how many images per hour you get on a short test run.
  2. Full 16-bit FLUX.1-dev without offloading, or several LoRAs and ControlNets at once: the RTX 4090 48G has 48 GB, more than the other two; the RTX 5090 is close to the limit for the full pipeline.
  3. Wan 2.2 14B video: more VRAM means less offloading, so start from a larger card, and check peak memory with nvidia-smi on your own resolution and frame count.
  4. Measure, then commit. Billing is per minute with a 1-minute minimum, so a short benchmark on two cards costs little. On /deploy, choosing a template shows only the GPUs that can run it.
# peak VRAM while one image or clip renders
nvidia-smi --query-gpu=name,memory.used,memory.total --format=csv -l 1

Frequently asked questions

Is the RTX 5090 faster than the RTX 4090 for image generation?

On paper, yes when the model fits on both: CloudGPU’s /pricing lists 1,792 GB/s and 209 FP16 TFLOPS for the RTX 5090 against 1,008 GB/s and 165 for the RTX 4090 (read 2026-11-03). Real speed depends on the model, precision and nodes, so time one image on your own workflow.

Is the RTX 4090 48G faster than the RTX 4090?

No. It is the same chip with 48 GB instead of 24 GB, so compute and bandwidth are the same. It helps when the 24 GB card has to offload or runs out of memory.

Can an RTX 4090 run FLUX.1-dev?

Yes with the FP8 checkpoint (about 12-16 GB of weights) or with fp8 weight_dtype or offloading. The full 16-bit pipeline is about 31.4 GB loaded in BF16, which is more than 24 GB.

Which card should I use for Wan 2.2?

The 5B model fits on all three; the ComfyUI tutorial says it “should fit well on 8GB vram” with native offloading. For the 14B models, more VRAM means less offloading; the Wan repo’s own script asks for at least 80 GB, so in ComfyUI use the FP8 files and measure on your settings.

Where are the prices?

On https://cloudgpu.app/pricing. This page does not quote prices.

Sources (read 2026-11-03, Asia/Shanghai)

The Deploy page lists the cards you can start. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-03 (Asia/Shanghai).