Replicate (qwen/qwen-image-edit-plus)

Editing an image used to mean a mask, an inpainting pass, maybe a ControlNet and an IP-Adapter stacked on top. Qwen Image Edit made a large slice of that obsolete: you hand it a picture and a sentence - "make her hair blue," "remove the lamp," "turn this into golden hour" - and the model just does it. This node is the ComfyUI front door to that model, running on Replicate's hosted API rather than on your GPU.

If the instruction-based editing workflow sounds familiar, it should: by 2026 this had become the default way the open-weights world edits images, and Qwen's edit line specifically won the category on licensing (Apache 2.0) as much as quality. The "Plus" variant is Replicate's boosted/optimized take on it, and that's what you're renting per prediction here.

How it works

This is one of the pack's auto-generated Replicate nodes, so the mechanics are shared with its siblings: the node sends your prompt and any image input to Replicate, polls the prediction to completion (streaming logs into the ComfyUI console), and downloads the result into an IMAGE tensor. You don't need any local model weights, VRAM, or CUDA - you need a Replicate account with a token and a willingness to pay per run.

The editing itself happens model-side: Qwen re-emits the whole frame according to your instruction. That's the power and the known catch - because the model regenerates everything, pixels you didn't ask about can shift, and a chain of edits can drift. The standard fix (mask the region you care about) still applies when you need surgical changes.

The inputs that matter

• prompt - the edit instruction. This is the whole game: "make the sky stormy," "turn the subject into an astronaut," "remove the background object." Be specific and terse; the model reads instructions well.
• image (optional IMAGE) - the image to edit. It's the only optional input, and it's what makes this an editing node rather than a text-to-image node.
• aspect_ratio - defaults to match_input_image, which is what you want 95% of the time for edits.
• go_fast - Replicate's speed-optimization toggle, on by default. Turn it off if you want maximum quality over speed; you'll pay for it in wall-clock time.
• seed - lock a seed to reproduce an edit, leave random to explore.
• output_format / output_quality - webp at 95 by default; the packed webp is a reasonable default for API results.
• disable_safety_checker - Replicate's safety filter toggle. On by default; flip it if the filter is rejecting legitimate edits.

Single output: IMAGE.

Install

Same pack as the whole family - ComfyUI Manager → search "Dados Nodes", or

cd ComfyUI/custom_nodes
git clone https://github.com/dadoirie/ComfyUI_Dados_Nodes.git
cd ComfyUI_Dados_Nodes
pip install -r requirements.txt

then restart. No local models to download.

The gotcha: your token

The pack's __init__.py hard-sets REPLICATE_API_TOKEN = "EMPTY" as a placeholder (with a "REMOVE LATER" comment above it), so this node - like every Replicate node in the pack - will fail auth until you set a real token in the environment that launches ComfyUI:

export REPLICATE_API_TOKEN=r8_...

Get it from replicate.com/account. People frequently report the node "doesn't work" when the real problem is this placeholder token; it's a five-minute fix that unblocks the whole pack's API family. And remember each prediction is billed to your Replicate account - instruction editing is great, but it's not free, so don't treat it as a replacement for local inpainting when you have the hardware.