萌宝AI·图片复刻设置

MengBaoImageReplicaSettings (萌宝AI·图片复刻设置) is the human-in-the-loop node of this pack's replication chain, and it's the one that decides whether the whole feature is worth using. Everything upstream is a machine guessing. This node is where you correct the guess - move boxes, rewrite copy, swap in real product photos, tell it to delete something - and it refuses to let you generate until you've said so out loud.

The chain it lives in

MengBaoLoadImage → MengBaoImageReverse → MengBaoImageReplicaSettings
                                              ├── prompt          → MengBaoImageAPI.prompt
                                              └── replica_settings → MengBaoImageAPI.replica_settings

analysis is the MENGBAO_REPLICA_ANALYSIS object produced by the reverse node. draft_json and draft_id are the node's own editing state - the panel writes your edits into them, and the node saves the draft back under that ID, so reopening the panel restores your work instead of re-reading the model's original guesses.

Outputs: prompt, a compiled STRING instruction set, and replica_settings, the structured payload the generation node actually consumes. Both go into the image node; the generation node then takes the original image as reference #1 and appends your replacement materials in element order, deduplicated by content ID.

What the panel actually does

Open it and the source image appears with numbered boxes over each detected element. For each one you choose keep or replace. For text elements you type the new copy. For non-text elements you describe the replacement, attach a reference image, or both. Move, scale, add or delete boxes as needed; positions are normalized, and after moving one you must press that item's Confirm position button, which is the kind of small friction that stops a half-dragged box from being silently used.

Deleting has real semantics, which is the part worth knowing: removing an element instructs the generator to erase it and repair the background behind it. Deleting a product also removes its packaging text and logo sub-items, because those belonged to it.

There's also an AI product description action that fires one extra vision request - only for product elements and their target reference images, not the whole source. Pressing it twice costs twice; the button disables while it's running.

When you're done, press Confirm settings. Any edit afterwards cancels that confirmation, and an unconfirmed node can't be submitted to the generator. Save draft explicitly does not authorise a generation either. This is a node designed by someone who got tired of paying for runs that turned out to be based on a stale edit.

Ratios and reference limits

The original is always reference image 1. The node picks the nearest aspect ratio that the selected generation model supports and appends a line to the prompt anchoring the canvas to it; the UI tells you when the source ratio isn't natively available. Reference ceilings include the source image - 14 for gpt-image-2, nano-banana-2 and nano-banana-2-pro; 16 for gpt-image-2.5 - and connecting ordinary reference image ports at the same time is an error, not a merge.

One quirk in the source's favour of predictability: in replica mode, gpt-image-2 and gpt-image-2.5 promote an auto resolution to a base 1K, because a fixed ratio can't coexist with fully automatic resolution on those backends. Your normal text-to-image runs are unaffected.

Installing it

Part of ComfyUI-MengBaoAI (Corkery520, MIT, Registry id mengbaoai, 17 nodes). Manager: MengBaoAI / mengbaoai / 萌宝AI. CLI: comfy node install mengbaoai. Git if the Registry listing isn't live yet:

cd ComfyUI/custom_nodes
git clone https://github.com/Corkery520/ComfyUI-MengBaoAI.git
cd ComfyUI-MengBaoAI
python -m pip install -r requirements.txt

numpy, Pillow, requests. This node's UI is JavaScript served by the pack, so after an update hit Ctrl+F5 - the README says exactly that, and stale front-end assets are the usual reason a panel button does nothing. Also disable the old standalone WANG plugins first.

Things that bite

• Position boxes describe placement, not real-world size. The docs are blunt: don't infer a product's true dimensions from a photo's crop or whitespace. Swapping in a photo whose subject is small in frame will scale it accordingly.
• Settings aren't self-contained in the workflow. The editing JSON and asset IDs are saved, but the assets live in ComfyUI/user/mengbaoai/replication/. Copy that directory when you move machines or you'll get missing-reference errors.
• Reading the original's text is still your job. Stylised and packaging lettering gets misread; the panel exists so you can fix it. Confirm before generating - it's cheaper than a re-roll.
• "Replica" is not a copy. Expect layout and mood to match, not pixels. If you need the exact original with one element changed, a local inpaint is the right tool, not this chain.