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

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. 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. LoRA training on the 12B transformer needs about 28.8 GB by the rule in our VRAM guide.
• Several models in one graph, high resolutions, FLUX.2 [klein] with everything loaded: L40S, 48 GB.
• FLUX.2 [dev] in BF16: one H200 (141 GB), or an 80 GB H100 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.

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.

# 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.