Best GPU for Stable Diffusion, Flux and AI video in 2026

The best GPU for Stable Diffusion and Flux, in short

The best GPU for Stable Diffusion and Flux is the cheapest one whose memory holds your model with room to spare. In our range that is usually the 32 GB RTX 5090 ($329 a month), which holds FLUX.1 [dev] in BF16 by our sizing rule and runs NVFP4 checkpoints on its Blackwell tensor cores. For 20–32B models in BF16, the next step is the 96 GB RTX PRO 6000 ($959).

• 16 GB, RTX 5080 ($199): SDXL, FLUX.2 [klein] 4B, FLUX 3 Image Klein, FLUX.1 in NVFP4.
• 24 GB, RTX 4090 ($269): SD 3.5 Large in BF16, FLUX.1 in FP8, Wan 2.2 5B.
• 32 GB, RTX 5090 ($329): FLUX.1 in BF16, Qwen-Image in FP8, FLUX.2 [dev] and LTX-2.5 in NVFP4.
• 48–96 GB, L40S ($789) or RTX PRO 6000 ($959): FLUX.2 [dev], Qwen-Image and LTX-2.5 in FP8 or BF16, with ECC memory.
• 80–141 GB, H100 ($1,779) or H200 ($2,279): video models split across several GPUs over NVLink, and large batches.

Prices are per GPU per month. Once a model fits, memory bandwidth and tensor throughput decide its speed. We publish no benchmarks: the figures below come from official model files and vendor specifications.

How much VRAM do Flux, SD 3.5, Qwen-Image and Wan 2.2 need?

A diffusion pipeline has text encoders for the prompt, a diffusion model that runs at every denoising step and a VAE that decodes the result into pixels. ComfyUI moves models between VRAM and system RAM as needed (README), so during sampling the GPU mainly holds the diffusion model and its activations, which grow with resolution, video length and batch size.

Parameter counts below come from the model cards, sizes from the file lists of the official repositories, checked on 19 October 2026.

Gated means you accept the conditions on Hugging Face with your own account before the files download. Wan 2.2 and HunyuanVideo 1.5 publish their weights at about 4 bytes per parameter; ComfyUI’s repackaged Wan 2.2 files are half that in FP16 and a quarter in FP8. The two A14B experts take turns (high-noise for the early steps, low-noise for the later ones), so 14B parameters work at each step.

Our sizing rule for generation counts 2.4 GB per billion parameters in BF16 and 1.2 GB in FP8, with 92% of GPU memory usable. For the diffusion model alone, that is 28.8 GB in BF16 or 14.4 GB in FP8 for FLUX.1, 48.0 GB or 24.0 GB for Qwen-Image, 76.8 GB or 38.4 GB for FLUX.2 [dev], and 33.6 GB or 16.8 GB per Wan 2.2 A14B expert.

FP8, NVFP4 and GGUF: how precision cuts memory

Precision is the main lever. BF16 stores 2 bytes per parameter and FP8 one. NVFP4, a 4-bit format introduced with Blackwell, adds one FP8 scale per 16 values; NVIDIA puts it at about 3.5× smaller than FP16 and 1.8× smaller than FP8 (NVIDIA). Real checkpoints shrink a little less: Black Forest Labs publishes FLUX.1 [dev] at 23.8 GB in BF16, 12.3 GB in FP8 and 9.2 GB in NVFP4, and FLUX.2 [dev] at 64.4 GB in BF16 and 21.0 GB in NVFP4.

• FP8 runs on every GPU in this article, as all have FP8 tensor cores. ComfyUI’s Flux tutorial points users with limited VRAM to the FP8 checkpoint, and ComfyUI publishes FP8 builds of Wan 2.2 (14.3 GB per expert) and of FLUX.2 [dev] (35.5 GB).
• NVFP4 is designed for the FP4 hardware of Blackwell GPUs, here the RTX 5080, RTX 5090 and RTX PRO 6000. On those cards ComfyUI reports about 2× faster sampling than with FP8 or BF16, but only with PyTorch built for CUDA 13.0; otherwise sampling can be up to 2× slower than FP8 (ComfyUI blog). Black Forest Labs’ NVFP4 files of FLUX 3 Image Klein need a Blackwell GPU (ComfyUI files). On an RTX 4090, L40S, H100 or H200, use FP8.
• GGUF quantizations can go below 8 bits through the community ComfyUI-GGUF custom node; its author notes that transformer models such as Flux suffer less from quantization than older UNet models.

Each step down can change the output, so compare results on your own prompts. When a model still does not fit, ComfyUI streams its weights from system RAM to the GPU. Its January 2026 update made that penalty smaller, not zero, and the weights need room in RAM: our servers have from 32 GB of RAM per GPU on the RTX 5080 to 180 GB on the RTX PRO 6000.

Best GPU for ComfyUI, card by card

The last column counts the diffusion model only; long videos and high resolutions need more.

RTX 5080: SDXL and the small new models

The RTX 5080 holds SDXL, FLUX.2 [klein] 4B, which Black Forest Labs says fits in about 13 GB, and FLUX 3 Image Klein, released on 15 October 2026, whose INT8 build runs on 8 GB cards (BFL). FLUX.1 [dev] fits as its 9.2 GB NVFP4 file. Bigger models mean offloading into 32 GB of RAM, and 0.5 TB of NVMe per GPU fills up quickly.

RTX 4090: SD 3.5 and FLUX in FP8

The RTX 4090 adds 8 GB: SD 3.5 Large in BF16 (19.4 GB by our rule), FLUX.1 in FP8 and HunyuanVideo 1.5 in BF16 (19.9 GB). Wan’s own scripts name it for the 5B model at 720p. It is an Ada card, with FP8 but no FP4.

RTX 5090: the default for Stable Diffusion and Flux

The RTX 5090 pairs 32 GB with 1.8 TB/s and FP4 tensor cores. FLUX.1 [dev] fits in BF16 (28.8 GB) while ComfyUI parks the T5 encoder in RAM, and Qwen-Image fits in FP8 (24.0 GB). FLUX.2 [dev] and LTX-2.5 run as NVFP4 checkpoints of 21.0 and 18.7 GB. See RTX 5090 vs RTX 4090.

L40S: 48 GB with ECC

The L40S holds FLUX.2 [dev] in FP8 (38.4 GB) or a Wan 2.2 expert in BF16 (33.6 GB), with ECC memory. But its 864 GB/s is about half the bandwidth of an RTX 5090, it has no FP4, and an RTX PRO 6000 costs $170 more for twice the memory. It suits many parallel workers that each need less than 48 GB, up to eight per server.

RTX PRO 6000: big models in BF16

The RTX PRO 6000 has 96 GB of ECC GDDR7 at 1.6 TB/s: Qwen-Image (48.0 GB), LTX-2.5 (52.8 GB) and FLUX.2 [dev] (76.8 GB) fit in BF16, or several models stay loaded at once. It meets the 80 GB that Wan 2.2’s single-GPU T2V-A14B command asks for, and MIG can split it into four 24 GB instances. Per GB it costs $9.99 a month, against $10.28 for the RTX 5090; see RTX PRO 6000 vs RTX 5090.

Best GPU for AI video generation

Video needs more activation memory than images, because all the frames of a clip are denoised together, so the official requirements are the better guide:

• Wan 2.2 TI2V-5B (720p at 24 fps): Wan’s single-GPU command needs at least 24 GB, such as an RTX 4090 (Wan 2.2). ComfyUI says the 5B model fits on 8 GB with its native offloading (ComfyUI docs).
• Wan 2.2 T2V-A14B: the official single-GPU command, with offloading and dtype conversion, needs at least 80 GB. With ComfyUI’s FP8 experts, our rule puts each at 16.8 GB, within an RTX 4090 or RTX 5090.
• HunyuanVideo 1.5: Tencent gives 14 GB as the minimum, with model offloading enabled.
• LTX-2.5, which generates video with audio: 42.0 GB in BF16 and 18.7 GB as the NVFP4 distilled checkpoint; Lightricks adds FP8 casting and CPU offload for smaller cards.

For one creator, an RTX 5090 runs these in FP8 or NVFP4, and an RTX PRO 6000 runs them in BF16 with room for longer clips. The H100 ($1,779) and H200 ($2,279) suit video at scale: Wan 2.2 and HunyuanVideo 1.5 ship scripts that split one video across 8 GPUs with torchrun, and in our H100 and H200 servers the GPUs exchange data over NVLink at 900 GB/s per GPU rather than PCIe. Their HBM, at 3.35 TB/s and 4.8 TB/s, feeds the cores faster, but neither has FP4.

Do two GPUs make one image faster?

Sometimes, but in ComfyUI they never add up to one bigger card: each GPU holds its own copy of the model (pull request), so two 24 GB cards do not hold a model that needs 40 GB. Since May 2026, ComfyUI’s MultiGPU CFG Split node shares one generation between identical GPUs by running the positive and negative prompt passes on different cards. ComfyUI has measured speed-ups of up to 1.95×, but only with CFG above 1, so step-distilled models that run at CFG 1, such as the distilled FLUX.2 [klein], gain nothing.

For more images or clips per hour, run one job per GPU: one ComfyUI instance per card, each started with its own --cuda-device and --port, fed from a queue. GeForce servers come with 1, 2 or 4 cards; L40S, RTX PRO 6000, H100 and H200 servers with up to 8, at the same price per GPU.

Where CryptGPU fits

CryptGPU rents dedicated GPU servers by the month, paid in crypto with no KYC, at about 30% below the median on-demand price of 40 GPU clouds. At the Software step of an order, ComfyUI can be preinstalled at no extra cost, with Qwen-Image, Wan 2.2 or SDXL downloaded during setup. FLUX and Stable Diffusion 3.5 require your own license acceptance on Hugging Face, so you download them after delivery with your own token. Our ComfyUI guide covers the setup and access through an SSH tunnel, and the image and video generation page lists recommended configurations.

Three limits to weigh. We publish no benchmarks: the figures above are file sizes and vendor specifications, not measured speeds. GeForce cards have no ECC memory; the L40S and RTX PRO 6000 do. And a monthly server pays off when you generate most days; for occasional use, an hourly cloud costs less. For language models on the same cards, see best GPU for LLM inference; to size a server and see its price, use the configurator.

FAQ

How much VRAM do I need for Flux?

FLUX.1 [dev] and [schnell] (12B) need 28.8 GB in BF16 by our rule, an RTX 5090, or 14.4 GB in FP8, an RTX 4090; ComfyUI can keep the 9.5 GB T5 encoder in system RAM. FLUX.2 [dev] (32B) needs 76.8 GB in BF16, or runs as a 21.0 GB NVFP4 file on Blackwell cards.

Is the RTX 5090 good for Stable Diffusion?

Yes. Its 32 GB holds SDXL, SD 3.5 Large and FLUX.1 in BF16 by our rule, and its FP4 tensor cores run NVFP4 checkpoints, which ComfyUI reports at about twice the speed of FP8 on RTX 50-series cards with PyTorch built for CUDA 13.0. It costs $329 a month.

How much VRAM does Wan 2.2 need?

The 5B model needs 24 GB with Wan’s own scripts, or about 8 GB in ComfyUI with offloading. T2V-A14B needs 80 GB with the official single-GPU command; ComfyUI’s workflow uses two 14.3 GB FP8 experts, one after the other.

Is 16 GB enough for ComfyUI?

For SDXL, SD 3.5 Large in FP8, FLUX.2 [klein] 4B, FLUX 3 Image Klein and FLUX.1 in NVFP4, yes. Larger models run with offloading, which is slower and limited by system RAM.

Which of these models allow commercial use?

Apache 2.0: FLUX.1 [schnell], FLUX.2 [klein] 4B, Qwen-Image and Wan 2.2. Free below a revenue threshold: SD 3.5 ($1M), FLUX 3 Image Klein ($5M) and LTX-2.5 ($10M). Non-commercial: FLUX.1 [dev], FLUX.2 [dev], FLUX.2 [klein] 9B and Qwen-Image-2.1. HunyuanVideo 1.5’s license excludes the EU, the UK and South Korea. Read each license before you build on a model.