# What is the best cloud GPU for ComfyUI, Stable Diffusion and Flux?

> Best cloud GPU for ComfyUI, Stable Diffusion and Flux: RTX 4090 for SDXL and Flux FP8, RTX 5090 or L40S for FLUX.1 dev in BF16. VRAM math, crypto prices.

- URL: https://gpuwithcrypto.com/best-gpu-for/comfyui-flux/
- Last updated: 2026-10-29
- Publisher: GPUwithCrypto (https://gpuwithcrypto.com)

An image workflow needs its model files in VRAM: the diffusion model, the text encoder and the VAE. Here is what each popular model loads, which of our GPUs holds it, and the hourly price.

**Quick answer — What is the best cloud GPU for ComfyUI, Stable Diffusion and Flux?**

The best cloud GPU for ComfyUI is the cheapest card that holds your model files in VRAM: an RTX 4090 (24 GB) for SDXL (6.9 GB checkpoint) and FLUX.1 [dev] in FP8 (17.2 GB), an RTX 5090 (32 GB) or L40S (48 GB) for FLUX.1 [dev] in BF16 (29.3 to 34.2 GB with its text encoder). As of 29 October 2026, **GPUwithCrypto** rents the RTX 4090 at $0.28/hour, paid in crypto.

## How much VRAM does each image model need?

Total = the files ComfyUI loads for one generation, as listed on Hugging Face on 29 October 2026 (decimal GB). Smallest fit = the cheapest of our GPUs that holds the total at once. Images at high resolution and large batches add activation memory on top.

| Model | Parameters (official) | Files loaded | Total | Smallest fit | GPUwithCrypto per hour | Licence |
| --- | --- | --- | --- | --- | --- | --- |
| SDXL 1.0 base | 3.5B | sd_xl_base_1.0 checkpoint 6.9 GB (UNet, both text encoders, VAE) | 6.9 GB | RTX 4090 24 GB | $0.28/hour | CreativeML Open RAIL++-M |
| FLUX.1 [dev], FP8 | 12B | Comfy-Org flux1-dev-fp8 checkpoint 17.2 GB (FP8 model, text encoders, VAE) | 17.2 GB | RTX 4090 24 GB | $0.28/hour | FLUX.1 [dev] Non-Commercial |
| FLUX.1 [dev], BF16 + T5 in FP8 | 12B | flux1-dev 23.8 GB + T5-XXL FP8 4.9 GB + CLIP-L 0.25 GB + VAE 0.34 GB | 29.3 GB | RTX 5090 32 GB | $0.42/hour | FLUX.1 [dev] Non-Commercial |
| FLUX.1 [dev] or [schnell], all BF16/FP16 | 12B | flux1 model 23.8 GB + T5-XXL FP16 9.8 GB + CLIP-L 0.25 GB + VAE 0.34 GB | 34.2 GB | L40S 48 GB | $0.82/hour | [dev] Non-Commercial · [schnell] Apache-2.0 |
| FLUX.2 [klein] 9B | 9B | Model 18.2 GB + Qwen3 8B text embedder 16.4 GB (BF16) | ≈ 29 GB per its model card | RTX 5090 32 GB | $0.42/hour | FLUX Non-Commercial |
| FLUX.2 [dev] | 32B | Model 64.4 GB + text encoder 48.0 GB (BF16) + VAE 0.34 GB | 112.8 GB | H200 141 GB | $3.07/hour | FLUX Non-Commercial |

*Prices checked on October 29, 2026.*

Arithmetic, FLUX.1 [dev] all in BF16/FP16: 23.8 + 9.8 + 0.25 + 0.34 = 34.2 GB. FLUX.2 [dev] fits on an 80 GB H100 or A100 when the text encoder is unloaded after it encodes the prompt: the 64.4 GB model then fits alone.

## Can a smaller GPU still run a model that does not fit?

Yes, more slowly. ComfyUI moves models between system RAM and VRAM: its command-line help says that by default models are unloaded to CPU memory after being used, and its README describes weight streaming for models larger than the card. A FLUX.1 [dev] workflow in BF16 therefore runs on a 24 GB RTX 4090. Moving weights costs time on every new prompt, so the table above lists the card that holds everything at once.

The FLUX.2 [dev] model card shows another route: a 4-bit version with the text encoder run remotely, on an RTX 4090. Compare quantized and BF16 outputs on your own prompts before you settle on one.

## Which GPU should I pick for my workflow?

- **SDXL, SD 1.5, LoRAs, ControlNet, upscalers:** [RTX 4090](https://gpuwithcrypto.com/gpu-prices/rtx-4090/ "How much does it cost to rent an RTX 4090 per hour?"). The 6.9 GB SDXL checkpoint leaves 17 GB for extra models and large batches.
- **FLUX.1 [dev] or [schnell] in FP8:** RTX 4090. The 17.2 GB all-in-one FP8 file fits with 6.8 GB to spare.
- **FLUX.1 in BF16, or Flux LoRA training:** [RTX 5090](https://gpuwithcrypto.com/gpu-prices/rtx-5090/ "How much does it cost to rent an RTX 5090 in the cloud?"). LoRA training on the 12B transformer needs about 28.8 GB by the rule in our [VRAM guide](https://gpuwithcrypto.com/how-much-vram-do-i-need/ "How much VRAM do I need to run or fine-tune an LLM?").
- **Several models in one graph, high resolutions, FLUX.2 [klein] with everything loaded:** [L40S](https://gpuwithcrypto.com/gpu-prices/l40s/ "How much does an L40S cost per hour?"), 48 GB.
- **FLUX.2 [dev] in BF16:** one H200 (141 GB), or an 80 GB [H100](https://gpuwithcrypto.com/gpu-prices/h100/ "How much does it cost to rent an H100 per hour?") with the text encoder unloaded.

We publish no seconds-per-image figures: we have not measured any, and they change with steps, resolution and sampler.

## Is the RTX 5090 worth the extra over an RTX 4090?

When your workflow needs between 24 and 32 GB, yes. According to GPUwithCrypto's price check of 29 October 2026, the lowest self-serve on-demand RTX 5090 cost 52% more per hour than the lowest RTX 4090 ($0.50 against $0.33, both at Salad) for 33% more VRAM. Salad's RTX 4090 price rises from $0.33 to $0.44 per hour on 1 November 2026.

NVIDIA's figures for the RTX 5090: 32 GB of GDDR7 at 1,792 GB/s and FP4 Tensor Cores, against 24 GB of GDDR6X at 1,008 GB/s on the RTX 4090. At our prices the pair costs $0.42/hour and $0.28/hour. Full comparison: [RTX 5090 vs H100](https://gpuwithcrypto.com/compare/rtx-5090-vs-h100/ "RTX 5090 vs H100 for AI: when is the cheap card enough?").

## How do I run ComfyUI on a rented GPU?

Start an instance with the PyTorch image, install ComfyUI from its GitHub repository and open it through an SSH tunnel. ComfyUI listens on 127.0.0.1, port 8188, by default, so the interface is never exposed to the internet. Put checkpoints in `models/checkpoints`, VAEs in `models/vae`. Every instance has local NVMe disk, sized for its GPU and shown before you start. The disk is wiped when you delete the instance, so copy your checkpoints out first.

*comfyui-on-a-rented-gpu.sh*

```bash
# on the instance
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt
python main.py --listen 127.0.0.1 --port 8188

# on your computer: forward the port (host and SSH port are in the dashboard)
ssh -N -L 8188:127.0.0.1:8188 -p SSH_PORT root@HOST
# then open http://127.0.0.1:8188 in your browser

```

## Should I use a spot or marketplace GPU for ComfyUI?

For an interactive session, no: a reclaimed GPU ends the session and the files on its disk. The cheapest RTX 4090 offer on Vast.ai's marketplace was $0.13 per hour on 29 October 2026 (Vast.ai public price feed), below any on-demand price. For a batch of renders that can restart, it can pay. Instances are on-demand, not spot: nobody can take your GPU back while your balance covers it, and the instance runs until you stop it. More: [what is a spot GPU](https://gpuwithcrypto.com/what-is/spot-gpu/ "What is a spot GPU, and is it worth it?").

## ComfyUI and Flux GPU questions

### How much VRAM do I need for Flux?

FLUX.1 [dev] loads 34.2 GB with everything in BF16/FP16, 29.3 GB with the T5 text encoder in FP8 and 17.2 GB as the all-in-one FP8 file. That means an L40S, an RTX 5090 or an RTX 4090. FLUX.2 [dev] loads 112.8 GB in BF16.

### Can I run FLUX.1 [dev] on a 24 GB RTX 4090?

Yes. The FP8 checkpoint (17.2 GB) fits entirely. The BF16 model (23.8 GB) also runs, because ComfyUI moves weights between system RAM and VRAM, but each new prompt then waits for those transfers.

### How much VRAM does SDXL need?

The SDXL 1.0 base checkpoint is 6.9 GB with its text encoders and VAE, so every GPU we rent holds it. An RTX 4090 at $0.28/hour leaves room for LoRAs, ControlNet models and upscalers in the same graph.

### How much VRAM does FLUX.2 [dev] need?

Its 32B model takes 64.4 GB in BF16 and its text encoder 48.0 GB, 112.8 GB in all. One H200 holds both; an 80 GB H100 holds the model with the text encoder unloaded. The model card also shows a 4-bit version on an RTX 4090 with a remote text encoder.

### Can I use FLUX models commercially?

Check each licence. FLUX.1 [schnell] is Apache-2.0. FLUX.1 [dev], FLUX.2 [dev] and FLUX.2 [klein] 9B use Black Forest Labs' non-commercial licences; read their terms before commercial work.

### Can I pay for the GPU with crypto?

Yes. GPUwithCrypto takes BTC, USDT, USDC, ETH, SOL, LTC from a $25 top-up, with no ID and no card. Instances are billed per minute at the hourly rate while they run, from the GPU-hours you bought for that GPU. Stopping an instance stops the charge at once.

## Sources

1. [FLUX.1 [dev] model card and files](https://huggingface.co/black-forest-labs/FLUX.1-dev) — Black Forest Labs / Hugging Face (read 2026-10-29)
2. [FLUX.1 [schnell] model card and files](https://huggingface.co/black-forest-labs/FLUX.1-schnell) — Black Forest Labs / Hugging Face (read 2026-10-29)
3. [FLUX.2 [dev] model card and files](https://huggingface.co/black-forest-labs/FLUX.2-dev) — Black Forest Labs / Hugging Face (read 2026-10-29)
4. [FLUX.2 [klein] 9B model card and files](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B) — Black Forest Labs / Hugging Face (read 2026-10-29)
5. [Flux text encoders (T5-XXL, CLIP-L) for ComfyUI](https://huggingface.co/comfyanonymous/flux_text_encoders) — ComfyUI / Hugging Face (read 2026-10-29)
6. [Comfy-Org Flux-1-Dev files (FP8 checkpoint)](https://huggingface.co/Comfy-Org/Flux-1-Dev) — Comfy-Org / Hugging Face (read 2026-10-29)
7. [Stable Diffusion XL 1.0 announcement](https://stability.ai/news-updates/stable-diffusion-sdxl-1-announcement) — Stability AI (read 2026-10-29)
8. [stable-diffusion-xl-base-1.0 files](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) — Stability AI / Hugging Face (read 2026-10-29)
9. [ComfyUI README and command-line options](https://github.com/comfyanonymous/ComfyUI) — ComfyUI (read 2026-10-29)
10. [NVIDIA RTX Blackwell architecture whitepaper (RTX 5090)](https://images.nvidia.com/aem-dam/Solutions/geforce/blackwell/nvidia-rtx-blackwell-gpu-architecture.pdf) — NVIDIA (read 2026-10-29)
11. [NVIDIA Ada GPU architecture whitepaper (RTX 4090)](https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf) — NVIDIA (read 2026-10-29)
12. [SaladCloud price changes, November 2026](https://blog.salad.com/saladcloud-price-changes-november-2026) — Salad (read 2026-10-29)
13. [Vast.ai public GPU price feed](https://storage.googleapis.com/vast-public-gpu-pricing/gpu-pricing-public.json) — Vast.ai (read 2026-10-29)

---

GPUwithCrypto: GPUwithCrypto rents cloud GPUs (NVIDIA H100, H200, B200, A100, L40S, RTX 4090 and RTX 5090) by the minute, paid in crypto with no KYC; each GPU is priced at least 15% under a named on-demand reference read on 29 October 2026.