RunPod SDXL (Custom Endpoint)

Like its sibling RunPodImageNodeSD, this is a client for your own RunPod serverless worker - specifically an SDXL one. Paste your endpoint ID and API key, dial in the prompt and sampler settings, and the node turns your canvas inputs into a job for your worker. It's the SDXL flavor of the pack's "custom serverless" trio, with the fuller param set that SDXL's two-stage generation deserves.

The pitch is the same as the SD node: you run ComfyUI somewhere weak but you've got a RunPod serverless SDXL deployment, or you want the pay-per-second-GPU economics without a local card. Where this node differs from its SD sibling is the refiner plumbing - it exposes refiner_inference_steps, high_noise_frac, strength, and a scheduler string, which is the shape of the classic base + refiner SDXL pipeline. If your worker template is the standard RunPod SDXL one, the field names line up.

How it works

The pack's shared async machinery with an endpoint input: POST to https://api.runpod.ai/v2/<your-endpoint-id>/run, poll until done, decode the image to an IMAGE tensor. The payload uses SDXL-style field names (num_inference_steps, guidance_scale, scheduler, high_noise_frac, refiner_inference_steps), so the worker on the other end needs to speak that dialect.

The inputs that matter

• api_key - RunPod key.
• endpoint - your deployed worker's endpoint ID. The field that makes this your node for your worker.
• prompt / negative_prompt - the generation text; SDXL does real CFG, so negative prompts matter.
• width / height - default 1024², SDXL's native range (up to 2048).
• steps - default 25; refiner_inference_steps - default 50, the second-stage count.
• guidance_scale - default 7.5.
• strength - default 0.3, and high_noise_frac - default 0.8: together they control how much of the denoise the refiner sees. The classic SDXL refiner split.
• scheduler - free-text, default DDIM. Match your worker's expectations.
• image_url - a URL for img2img when you want to run from an image rather than pure text.
• seed - -1 random; set it to reproduce.

What comes out

One image output (IMAGE). Save, preview, chain an upscaler - normal tensor.

Installing it

One of the four custom nodes in tcarwash/ComfyUI_RunpodNodes. ComfyUI Manager: search ComfyUI_RunpodNodes, or

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

then restart. No extra dependencies - the pack only needs what ComfyUI already ships. You need a RunPod key and a deployed SDXL serverless endpoint.

Gotchas

• The worker is the contract - the node sends a specific payload shape; if your worker template differs (different template, custom script), fields silently no-op or the job fails. Test with the worker's own docs first.
• Refiner fields only do something if your worker runs a refiner - strength and high_noise_frac are the base/refiner split; a single-model worker ignores the intent.
• Serverless metering - per-second GPU billing, cold starts, idle costs. Keep the worker warm only if you actually use it.
• scheduler and sampler_name are free text across this pack - no dropdowns to save you from a typo.

The bottom line: if you've gone through the work of deploying an SDXL worker, this node is the cleanest way to drive it from ComfyUI. If you haven't, the public-endpoint image nodes in this pack get you an image with a fraction of the setup - SDXL serverless is for people who already made that choice.