Qwen Image Edit Plus (Replicate)

"Make the sky stormy," "turn this into a watercolor," "remove the sunglasses." Qwen Image Edit Plus is the edit model from the Qwen family, and what it understands is instructions, not inpainting masks. You hand it an image and a plain-English sentence about what changed, and it rebuilds the image to match. It's the closest thing in this pack to a Photoshop "generative fill" that doesn't need you to paint a mask first.

The reason you'd route this through a node instead of the Qwen website is the same as every other member of Replicate Select: the model is closed, hosted, per-call - but it lives inside your workflow. Feed it the output of a local generator, ask for a change, wire the result onward. No download, no VRAM, and the editing loop becomes part of the graph instead of a copy-paste detour to a website.

How it works

Catalog entry qwen/qwen-image-edit-plus, handled by the pack's standard engine. One important detail from the source: its image input is flagged as a list, so if you wire in a batch the pack uploads each frame individually rather than cramming them into one request. Your reference image goes up as PNG, the prompt rides along, and the edited result comes back as an IMAGE tensor.

The inputs that matter

• image (required) - the image to edit. The tooltip is specific: must be jpeg, png, gif, or webp. You're feeding IMAGE tensors, so you're fine, but don't point this at anything exotic.
• prompt (required) - the instruction. This is where the model earns its keep: "add a lighthouse on the cliff" works far better than a negative-prompt wall of words. Describe the change, not the whole image.
• aspect_ratio - defaults to match_input_image, which is what you want for edits that shouldn't change composition.
• seed - set it when you want the same edit reproducible. 0 means "surprise me."
• go_fast - on by default, additional optimizations for faster predictions. Leave it on unless you notice quality loss.
• output_quality - 0–100, default 95, ignored for png outputs.
• output_format - webp default, or jpg/png. If you're going to keep editing, png avoids generational loss.
• disable_safety_checker - off by default, and same story as the rest of this pack: it relaxes, it doesn't remove.

Single output: image, an IMAGE tensor, ready to plug into SaveImage or another node.

Install

This node is one member of the Replicate Select pack - install the pack, get all of them:

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", then restart. Dependency: replicate>=1.0.7 only. Auth via environment variable, not a node field:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• It's a re-render, not a touch-up. Qwen-Edit rewrites the image to satisfy the instruction. Small, surgical changes (a birthmark, a specific logo) can drift; for those, a masked inpainting workflow on a local model is still the better tool. Instruction editing excels at whole-scene or style changes.
• The batch surprise. Because the image input is treated as a list, a multi-image batch means multiple uploads and multiple charges - check what you're feeding in before you queue a big run.
• Composition drift. Without match_input_image (or with a strong instruction), the model can change framing. If layout matters, keep that default and phrase the prompt to preserve it.