RunPod Qwen Image Edit

Qwen-Image-Edit is the model that made a giant class of image work stop needing a mask. You hand it a picture and a plain-language instruction - "change the color to light grey" is literally the node's default prompt - and the change happens. Clothing swaps, background relights, object removal, all without a ControlNet in sight. It won its category over Flux Kontext mostly on licensing, and by 2026 it's the default instruction editor people point to. This node is that model (qwen-image-edit) as a RunPod public endpoint, so the editing skill is one node away and your GPU never has to lift a finger.

The "without a mask" framing matters more than it sounds. Editing used to mean: mask the region, pick an inpaint model, fiddle with a control map, pray. Qwen-Image-Edit collapses that whole rig into a sentence. It's a 20B model doing it, which is exactly why the hosted route makes sense - you'd rather rent the 20B per edit than live with it.

How it works

Standard pack mechanics: async job to RunPod, polling in the console, base64 back to an IMAGE tensor. The node-specific bit is the image input - you have both an optional input_image wire (from a Load Image node, the one you'll actually use) and an image URL text field, and the node converts your tensor into a data URL the endpoint accepts. The output is your edited image, ready to save or continue through the graph.

The inputs that matter

• input_image - the source picture. This is the whole job.
• prompt - the instruction. Say what to change and, helpfully, what to keep. Vague instructions get vague edits.
• negative_prompt - Qwen has real CFG, so this earns its keep here more than on the Flux nodes.
• seed - -1 random; set it to reproduce an edit you liked.
• output_format - png (default) or jpeg.
• enable_safety_checker - on by default; the endpoint enforces its own policy and there's no bypass.

What comes out

One image output (IMAGE). Save it, upscale it, or chain more edits - the output is a first-class tensor like any local generator's.

Installing it

One install covers the whole 27-node pack. ComfyUI Manager: search ComfyUI_RunpodNodes, or

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

then restart. No extra dependencies - requests, Pillow, torch are already there. You need a RunPod API key from runpod.io.

Gotchas

• It's metered - the hosted Qwen Image Edit endpoint is a real, current RunPod public endpoint (a RunPod staffer has pointed people at it directly when local options were blocked), but that means per-image billing and queue waits. For heavy editing sessions the cost adds up.
• 30–120 second "hang" - it's polling. Watch the console.
• Filtering at the source - the safety checker is on and the model refuses what it refuses; there's no local weights to patch.
• "Job failed" - key or credits, the pack's eternal suspects.

If you already run Qwen-Image-Edit locally (it's Apache 2.0, so you can), local wins on cost and privacy. This node is for the no-GPU machine, the busy GPU, or the "I just need this one edit done right now" moment. For any of those, it's about the most direct editing node in the pack.