RunPod Qwen Image

Qwen-Image is Alibaba's 20B multimodal diffusion transformer, and its one genuinely untouchable skill is text rendering - real typography, in English and Chinese, that survives inside an image without turning into letter soup. Almost nobody runs the 20B base as a daily driver locally, because almost nobody has a card that wants 20B around all day. This node is the sensible middle path: it calls Alibaba's qwen-image-t2i model through RunPod's public endpoint, so you get the text-rendering flagship per image, on a bill, instead of permanently reserving VRAM for it.

How it works

The standard pack plumbing: your prompt and settings go to RunPod as an async job, the node polls until it completes, and the returned base64 decodes into a ComfyUI IMAGE tensor. Nothing runs locally except the HTTP client - which is the entire point. The endpoint handles the 20B model, the VAE, and the Qwen2.5-VL condition encoder that makes the family so good with language.

The inputs that matter

• api_key - RunPod key, shared via a Primitive if you have several nodes from this pack in one graph.
• prompt - the image description. For text rendering, write the actual text you want in the image and let the model lay it out; that's the party trick.
• negative_prompt - sent along, and unlike Flux, Qwen has real CFG so this one does something. Worth filling in.
• size - a string, default 1024*1024. Keep the W*H shape.
• seed - -1 random; set it to reproduce a layout you liked.
• enable_safety_checker - on by default, refuses at the source.

What comes out

One image output (IMAGE). Preview it, save it, or feed it into an upscaler - a text-heavy image is exactly the thing you'll want at higher resolution, so don't be shy about chaining.

Installing it

It's one of the 27 nodes in tcarwash/ComfyUI_RunpodNodes. Install via ComfyUI Manager (search ComfyUI_RunpodNodes), or:

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

Restart. No requirements.txt, nothing heavy - the pack runs on requests, Pillow, and torch, all already in ComfyUI. You'll need a RunPod key from runpod.io.

Gotchas

• It's a convenience, not a capability unlock - Qwen-Image's weights are Apache 2.0, so if you have the VRAM, local beats metered. The node wins when you don't.
• Queue + polling - public endpoints take 30–120 seconds; the console's [RunPod] Status: lines are your only feedback, so keep the terminal visible.
• Per-call cost - a 1024² image is roughly a megapixel of billing. Fine for real use, annoying if you're doing 200-prompt grid searches.
• "Job failed" - key or credits. The pack's two classic suspects.

One more honest note: for text-heavy work, Qwen-Image is genuinely the one to reach for. If you need editing with that same text skill, skip this node and grab its sibling RunPod_QwenImageEdit instead - the base model gets admired, the editor gets used.