RunPod Qwen Image LoRA

This is the Qwen Image base model plus a LoRA hook, as a RunPod public endpoint. Same 20B text-rendering powerhouse as RunPod_QwenImage, but with an extra loras input that takes a JSON array so you can pin a style or a character onto a model you never actually host. It's the node for "I want Qwen's typography and my LoRA's look," without a 20B model squatting in your VRAM.

If you've used LoRAs locally this is familiar territory - the difference is where the weights live. Here the LoRA JSON rides in the request payload and the endpoint's worker loads it server-side. You're buying the convenience of Qwen's base plus a LoRA stack at per-image prices, which is either exactly what you need or overkill, depending on how much iteration you plan.

How it works

Standard pack plumbing: async POST to qwen-image-t2i-lora, polling, base64 image back out. The node-specific input is the loras field, which the code parses as JSON (it's one of the pack's designated JSON-array params, along with images and the WAN high/low_noise_loras fields). Give it a well-formed array and the endpoint applies them; malformed JSON and you get a clear Invalid loras: error instead of silent weirdness, which is a small mercy.

The inputs that matter

• api_key - RunPod key, shared via Primitive.
• prompt - the generation text. Qwen's strength means you can include real text in the image.
• loras - a JSON array, e.g. [{"model": "sdxl/characters/your-lora", "scale": 0.8}]. The exact schema depends on what the endpoint expects, so check RunPod's docs for the current format before you fight it.
• size - string, default 1024*1024.
• seed - -1 random; set it to reproduce.
• enable_safety_checker - on by default.

What comes out

One image output (IMAGE). Preview, save, upscale - it's a normal tensor.

Installing it

One of 27 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 requirements.txt drama - the pack needs only requests, Pillow, torch. Grab a RunPod key at runpod.io.

Gotchas

• The LoRA schema is the fiddly part - it's a hosted endpoint, so your LoRA has to be reachable by the server (usually via a URL or a model ID the endpoint knows), and the JSON shape has to match. This is the node in the pack where you'll spend time in the API docs, not the node UI.
• Queue + metered cost - 30–120 seconds of polling and per-image billing, same as the rest of the family.
• "Job failed" - check key, credits, and then your LoRA JSON. In that order.

Honestly: if you need LoRAs, you probably have a GPU or access to a local setup, because LoRA culture is a local thing. This node is for the no-GPU case or the "render once with a LoRA I want to try" case. When that's your situation, it's a genuinely clean path - just budget time for the JSON.