Stable Diffusion Hosting: ComfyUI on a Dedicated GPU

Stable Diffusion hosting has a resource problem that text-model hosting doesn't: the workflows people actually want to run — SDXL with a refiner, ControlNet, a stack of LoRAs, or Flux — chew through VRAM fast, and consumer cards keep you permanently one node short. SnapDeploy's one-click ComfyUI template puts the most popular Stable Diffusion interface on your own dedicated NVIDIA GPU — with the 24 GB A10G tier ($999/month flat) as the comfortable home for SDXL-class work — running 24/7 at a public HTTPS URL. Here's what fits where, how the template works, and when flat-rate beats per-second for image generation.

What Is ComfyUI?

ComfyUI is a node-based interface for Stable Diffusion: instead of a fixed prompt box, you wire up a graph — checkpoint loader, prompt encoders, sampler, VAE decode, upscaler — and every graph is a shareable JSON workflow. That graph model is why it has become the power-user standard: community workflows for SDXL pipelines, LoRA stacking, ControlNet posing, and inpainting can be dragged into the canvas and reproduced exactly. It serves a web UI on port 8188, plus an HTTP API for queueing prompts programmatically — which is how a hosted ComfyUI doubles as the image-generation backend for a product, not just a canvas for a person.

VRAM Reality Check: Why SDXL Wants 24 GB

Model cards list minimums; workflows have appetites. What single-GPU image generation actually looks like:

SDXL can run in 12-16 GB with memory offloading, and ComfyUI is famously good at squeezing models into small cards. But "runs" and "works well" are different products: on a cramped card you get batch-of-one renders, slowdowns as models shuffle between VRAM and system RAM, and out-of-memory errors the moment you add a ControlNet. On the A10G's 24 GB, SDXL at full 1024x1024 resolution keeps the base model, refiner, and extras resident at once. That's why the ComfyUI template recommends the A10G tier ($999/month, ₹96,903 in India) — the T4 ($499/month) remains a fine choice for SD 1.5-era checkpoints.

How the Template Works

The one-click template — open source at github.com/AAR-Labs/comfyui-gpu-snapdeploy — is built around a background SDXL download. The ComfyUI server starts and passes health checks immediately; the multi-gigabyte SDXL checkpoint downloads behind it on first start. Your URL is live in 2-3 minutes, and the default SDXL workflow becomes runnable as soon as the checkpoint lands — no SSH session, no wget-ing model files by hand, no CUDA setup.

• Workflows — drag any community workflow JSON into the canvas, or queue prompts programmatically through the API.
• LoRAs and custom checkpoints — fork the template repo and add download steps for the models you want available at startup. LoRA files are small (tens to hundreds of MB) and stack on the base checkpoint in your workflow.
• Flux-schnell — the fp8 quantized variants fit in the A10G's 24 GB; add the model downloads to your fork to run Flux workflows. Full-precision Flux checkpoints are borderline at 24 GB — stick to fp8.

Step-by-Step: Deploy ComfyUI on SnapDeploy

1. Create a free SnapDeploy account and subscribe to Dedicated GPU from the billing page — A10G recommended for SDXL and Flux. See full pricing.
2. Click "New Container" and select GPU as the compute type.
3. Pick the ComfyUI template from the template gallery.
4. Click Deploy — the build takes 2-3 minutes, then the UI is live on port 8188 at your public HTTPS URL while SDXL finishes downloading in the background.
5. Open the URL, load a workflow, queue a prompt. ComfyUI ships no authentication, so treat the URL as a secret — or fork the template and add an authenticating proxy in front.

Flat-Rate vs Per-Second for Image Work

Image generation is an iterative workload — you don't render one image, you render forty variations while tuning a prompt, then come back after lunch and render forty more. That shape is exactly where per-second billing gets awkward:

• The meter runs while you think. A per-second GPU bills for the minutes you spend staring at a render deciding what to change — or you scale to zero between iterations and eat a cold start each time.
• Cold starts are heavy here. Re-loading an SDXL or Flux checkpoint after scale-to-zero takes minutes, not milliseconds. A dedicated GPU keeps checkpoints warm in VRAM 24/7 — the queue starts sampling immediately, every time.
• The bill cannot spike. A10G is $999/month flat (~$1.37/hr effective over ~730 hours) whether you render ten images or ten thousand. A busy launch week changes nothing.

The honest caveat, same as ever: if you generate one batch a week, a per-second provider is cheaper for that burst. Flat-rate wins when image generation is part of your working day, your team's pipeline, or your product's backend — and since the GPU is dedicated, one subscription can serve a whole design team hitting the same ComfyUI instance. Full provider-by-provider numbers in our GPU cloud pricing guide.

Frequently Asked Questions

How much VRAM does Stable Diffusion XL need?

SDXL can technically run in 12-16 GB with memory offloading, but it's cramped — batches of one, slowdowns from model juggling, and no headroom for the refiner, ControlNet, or LoRA stacks. On a 24 GB A10G, SDXL at 1024x1024 runs comfortably with room for extras. That's why the ComfyUI template recommends the A10G tier.

Can I use my own checkpoints and LoRAs?

Yes. The template is an open-source GitHub repository — fork it and add download steps for any checkpoints, LoRAs, or ControlNet models you want available at startup, then deploy your fork. LoRA files load on top of the base checkpoint inside your workflow.

Does Flux run on SnapDeploy?

Flux-schnell in its quantized fp8 form fits within the A10G's 24 GB of VRAM and runs in ComfyUI. Full-precision Flux checkpoints are borderline at 24 GB — expect to use the fp8 variants. Fork the template and add the Flux model downloads to try it.

Is the ComfyUI interface public?

Your container gets a public HTTPS URL, and ComfyUI does not ship with built-in authentication. Treat the URL as a secret, or fork the template and put an authenticating reverse proxy in front before sharing access beyond your team.

Why not use a per-second GPU service for image generation?

For occasional one-off batches, a per-second provider is genuinely cheaper — see the economics section above. Flat-rate wins when image generation is part of your working day or product: no meter running while you iterate, no multi-minute checkpoint reloads after cold starts, and a bill that stays $999/month no matter how many images you render.

Getting Started

Hardware details, tier comparison, and the full platform FAQ are on the GPU hosting page. The short version:

1. Create a free SnapDeploy account
2. Subscribe to Dedicated GPU — A10G for SDXL and Flux, T4 for SD 1.5-class checkpoints
3. Deploy the ComfyUI template and open your public URL
4. Load a workflow and start rendering — no meter, no cold starts, your own GPU

Serving language models instead? See our companion guide to Ollama hosting on a dedicated GPU.