▶️ RH Execute

RH Execute is the node that actually makes the cloud run. You connect a config from RH Config, hit Queue Prompt, and this node creates a task on RunningHub, waits for it to finish on their GPU, downloads the results, and hands them back to your graph as if a local sampler had produced them. It's an API-wrapper node in the classic sense - from the canvas it looks like a generator, but underneath it's an HTTP client spending credits on someone else's hardware.

The reason to reach for it: you want a RunningHub workflow inside your local ComfyUI, with your upscaler, masking, or post-processing still running locally around it. Or you have no GPU worth the name and this is your access to one. If the "model is on a server I don't control" part bothers you, this whole pack is the wrong tool - the tradeoffs (per-call cost, inputs leaving your machine) are spelled out in the KB's external-api-nodes essay, and they're real.

How it works

Under the hood it's three steps. First it POSTs to /task/openapi/create (or /task/openapi/ai-app/run for AI Apps) with your workflow ID and the nodeInfoList of parameters, with five retries and exponential backoff. Then it polls /task/openapi/outputs every five seconds until the task finishes - a busy wait, so the node blocks your queue until done. Finally it downloads the output files in parallel (up to 10 workers) and converts them: PNGs/JPGs to image tensors, videos to frame tensors via OpenCV (plus the raw file saved if you asked), text, audio via torchaudio, and .safetensors latents.

It even prints a direct link to https://www.runninghub.cn/task/detail/<task_id> in the console, which is genuinely useful mid-run.

Inputs that matter

• config - from RH Config. Required.
• timeout - default 600s, range 60–3600. This is the max wait, and "Task timeout" is the #1 support question. Cloud queues back up; give it room.
• params - the optional RH_PARAMS chain from RH Param or the upload nodes. Skip it and the workflow just runs with its saved defaults.
• use_high_performance - flips to a 4090 48GB instance (instanceType: plus). Costs more credits. Only worth it when the default queue is slow or the workflow needs the VRAM.
• save_to_local / output_prefix - write results into your ComfyUI output folder with a prefix like RH_20260826_..._001.png.

Outputs

Six result outputs - images, video_frames, text, audio, video, latent - plus task_id. The first thing you'll do is wire images into a Preview Image node. When a category comes back empty you don't get nothing; you get placeholder outputs (a "No images" placeholder image, empty text, a silent audio tensor), which keeps the graph alive but can look like silent failure if you're not expecting it.

Keep the task_id. It's the key to RH Download (fetch results later, even in a fresh session) and RH Task Manager (check status, cancel).

Install & gotchas

Via ComfyUI Manager (search ComfyUI_RH_API) or git clone https://github.com/xuchenxu168/ComfyUI_RH_API.git into custom_nodes, then pip install -r requirements.txt, restart. No model downloads - they're on the cloud.

Watch the business errors: INVALID_API_KEY, INSUFFICIENT_BALANCE, and especially WORKFLOW_NOT_SAVED_OR_NOT_RUNNING - that last one means the workflow on RunningHub is still in Draft or the ID is stale. And remember RunningHub moderates content, so anything it refuses never comes back. Finally, this node blocks: long workflows with big timeouts will hold your local queue hostage, which is exactly why the pack splits out batch runs and result download into separate nodes.