萌宝AI·复刻检查

Cloud image models are excellent at rendering text and completely untrustworthy at rendering the right text. You asked for the new slogan "SOFT AS CLOUD" on eight images and six came back with something close enough to look fine at thumbnail size - plus one where the old price survived in the bottom corner because the original reference bled through. Nobody catches that at 2 a.m. on the last page of a set.

MengBaoReplicaAudit (萌宝AI·复刻检查) is the QA node for that. It's the last link in the replication chain and it's the only one whose job is to disagree with the generator.

What it checks

Wire images (the generated batch) and replica_settings (from MengBaoImageReplicaSettings) into it. For every image in the batch, it asks a vision model whether the target copy is present and correct, whether any text has been altered or garbled, and whether old copy from the original that you replaced or deleted is still visible. The results come back as a JSON report on the report string output, with one entry per image carrying the model used, how many attempts it took, and whatever the model found.

The images output is a pass-through of your input, so you can slot the audit between the generator and your save node without restructuring anything: generator → audit → save. That's the intended placement, and it means the audit is also your visual confirmation step before files hit disk.

Failure behaviour, which is the good part

• One vision call per image. Four generated images, four billable requests. Budget accordingly - this is a per-image tax on your batch.
• Per-image failures don't kill the batch. If one audit call errors, that image's report entry records the error and the loop continues to the next. The images stay in the output either way; the README is explicit that a failed check never discards the generated artwork.
• Cancelling stops the remaining checks, and the images you already have still pass through.
• A failed audit does not touch generation history. It can't mark a generation as incomplete or trigger a re-roll, and it never re-bills the generation itself. The audit and the generator are separate meters.

The same vision model stack as the reverse node is used - gem-3.7-flash with a single fallback to gem-3.8-flash, 180-second timeouts, switching only on connection errors, timeouts, 429, 5xx and a clearly unavailable model, and using only the pack's global API key. A key pasted on the image node doesn't apply here.

Reading the report

report is a JSON array pretty-printed with non-ASCII preserved, so Chinese copy stays readable. Each element is keyed by image number, which lines up with batch index. In practice you skim it: the useful signal is "image 5 says the old tagline is still visible", not the full prose the model returns. Wire it into a text display node next to the save node and it's visible on the canvas without digging through a file.

One honest limitation the docs state out loud: uncertain packaging text and artistic lettering still need a human. The audit narrows the haystack; it doesn't replace looking at the images.

Installing it

It ships in ComfyUI-MengBaoAI - 17 nodes, Corkery520, MIT, Registry id mengbaoai. Manager: search MengBaoAI, mengbaoai or 萌宝AI. CLI: comfy node install mengbaoai. Git while the Registry listing is still clearing:

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 - no model downloads, but the vision calls need the global API key saved first, or you get Global API Key is empty. Save a global API Key first. Restart, then search 复刻检查, Replica Audit, or 文字检查. If you're coming off the standalone WANG plugins, disable them before this pack loads.

Things that bite

• Cost scales with batch size. Eight generated images is eight audit requests on top of eight generation requests.
• It audits against replica_settings, not your eyes. If you skipped the Confirm step, or edited copy afterwards without re-confirming, you're auditing the wrong target text.
• Vision models misread stylised type too. A false "missing text" on a fancy script headline is normal. Treat the report as a ranked suspicion list, not a verdict.
• Not a general-purpose caption or safety checker. It compares copy; it won't tell you the product looks wrong.