11 November 2026 · VRAMfarm team

Best GPU for Flux, SDXL and ComfyUI: 16 GB vs RTX 3090, 4090, 5090 and 48 GB cards

If you mainly run SDXL, a 12–16 GB card is enough. For Flux.1 dev, 16 GB works and 24 GB is comfortable. Qwen-Image and HiDream want 24 GB, and Wan 2.2 14B video wants 32 GB at the very least and 48 GB to work without fighting the memory limit. Within one memory size, the faster card is often no more expensive per image, because it finishes sooner.

Memory first: what each model needs in ComfyUI

ComfyUI is good at squeezing models into less memory. It loads the text encoder, runs it, moves it out, then loads the diffusion model. So the number that matters is mostly the largest single model in the workflow, plus room for the image itself.

The VRAM figures are estimates for a plain workflow at 1024 × 1024 or 720p; LoRAs, ControlNet, resolution and clip length push them up. The “download” column matters for your bill. The ready-made apps fetch their models after the server starts, and that traffic is billed at the rate the machine shows before launch.

A note on FLUX.2 [dev], the 32B successor to Flux.1: its fp8 weights alone are around 32 GB, and its text encoder is a large language model in its own right. It runs on 24–32 GB cards only with heavy offloading to system RAM. Treat 48 GB as the comfortable size.

Formats matter too. A full bf16 Flux.1 dev checkpoint is about 24 GB, which only just fits on a 24 GB card. The fp8 version halves that with results that are hard to tell apart. GGUF quantisations (Q4, Q5, Q8) shrink it further and let Flux run on 8–12 GB, at the cost of slower steps and slightly softer detail.

When 16 GB is enough

A 16 GB card (RTX 4060 Ti 16 GB, RTX 4080, RTX 5070 Ti, RTX 5080, RTX A4000) covers more than people expect:

• SDXL with LoRAs, ControlNet and an upscaler, at 1024 × 1024, in batches of 2–4.
• Flux.1 dev in fp8 at 1024 × 1024, one image at a time, with a LoRA or two.
• Wan 2.2 5B at 480p, or at 720p with offloading (slower).
• Qwen-Image in GGUF Q4, if you accept some loss of detail.

16 GB starts to hurt when you stack things: Flux plus two ControlNets plus a LoRA at 1536 × 1536, Qwen-Image in fp8, any training, or any serious video. You will see out-of-memory errors or, worse, ComfyUI silently offloading to system RAM and each step taking several times longer.

If your work is SDXL or a few Flux images a day, the cheapest card that holds the model is the right choice. On a 16 GB card you pay from €0.16 an hour for the 4060 Ti or €0.31 for the 4080. Between the two, the 4080 has more than twice the memory bandwidth (717 vs 288 GB/s), and image generation feels it.

24 GB: RTX 3090 vs RTX 4090

The two cards hold the same 24 GB, so they run exactly the same models. The difference is speed.

In diffusion models, which are mostly compute-bound, the RTX 4090 usually finishes an image about 1.6–2.5 times faster than the 3090. The top of that range applies to fp8 models when the software uses the 4090’s native fp8 maths (in ComfyUI, the --fast start option); the 3090 always converts fp8 weights to 16-bit to compute with them. Exact times depend on steps, resolution, sampler and how many nodes your workflow has, so treat the ratio as a guide, not a benchmark.

To decide, compare the price ratio with the speed ratio. Cost per image is the price per hour × seconds per image ÷ 3,600. If the RTX 4090 costs less than about twice as much per hour as the RTX 3090, it is usually also cheaper per image. If it costs more than twice as much, the 3090 usually wins on money and the 4090 wins on your time. Check both prices on the pricing page at the moment you rent, because the catalogue changes during the day.

The rule of thumb that falls out of this: for an interactive session where you tweak prompts and wait for each result, take the 4090. For a long overnight batch where nobody waits, the 3090 is often the better buy per image.

RTX 5090 and 48 GB cards

The RTX 5090 adds two things: 32 GB instead of 24, and 1,792 GB/s of memory bandwidth on top of more compute. For image models that means roughly 1.3–1.5 times the speed of a 4090, more where software uses Blackwell’s fp4 support, and room to keep Qwen-Image or HiDream fully loaded with LoRAs instead of shuffling models in and out. It is also the smallest card that runs Wan 2.2 14B in fp8 at 720p.

48 GB cards are a different bet: more memory, not more speed.

Take a 48 GB card when the model does not fit in 32 GB: Wan 2.2 14B at 720p with longer clips, Qwen-Image in bf16, FLUX.2 without offloading, or LoRA training with large batches. For plain Flux.1 or SDXL, a 48 GB card mostly costs more per hour for memory you don’t use. An RTX 4090 is usually faster than an RTX A6000 on the same Flux image.

Video: Wan 2.2 5B and 14B

Video is where memory decides everything, because a 5-second clip is 80–120 frames processed together.

• Wan 2.2 5B generates 720p video on a 24 GB card. It is fast enough to try many prompts, and it is the right model for drafts. Rent an RTX 4090 or RTX 5090.
• Wan 2.2 14B gives noticeably better motion and prompt-following. In fp8 it needs about 30 GB at 720p for a 5-second clip. A 32 GB RTX 5090 can do it; a 48 GB L40S or RTX A6000 does it with room to spare. Our ready-made Wan 2.2 14B app only offers 48 GB cards or larger, because short clips at 720p are the minimum people want from it.

A clip takes minutes, not seconds. Speed LoRAs that cut the number of steps (4–8 instead of 20–40) reduce that a lot, with some loss of detail. Plan for 10–15 minutes after the server is ready before you can start, because the 14B app downloads about 70 GB of models. More on the full pipeline is on the video generation page.

Batch size, speed and what to rent

More VRAM lets you raise the batch size, but don’t expect it to multiply your output. A big diffusion model at 1024 × 1024 already keeps the GPU busy with a batch of one, so a batch of four takes almost four times as long and often gains you only 10–30% in throughput. Batching helps more with small models like SDXL, where one image does not fill the card. Its real value is convenience: you queue four variants and look at them together.

A short decision list:

If you are not sure your workflow fits, start one size up for the first session and watch the memory in nvidia-smi or the ComfyUI console. You pay per minute, so a test run costs little. The VRAM guide has the numbers for LLMs, speech and training, and our ComfyUI guide covers adding your own models, LoRAs and nodes.