Meta Apply

Here's the killer workflow in this pack, and it's one idea: drag a PNG you made last week into this node and the entire recipe that produced it - checkpoint, LoRAs, sampler settings, prompts - gets pushed back into the current workflow. No transcription, no "which seed was it again," no digging the .json out of your ComfyUI folder. The metadata was riding inside the file all along; this node just applies it.

How it works

Meta Apply is a UI-only output node: it has no data outputs and does no computation of its own at execution time. What it does is sit on the canvas as a drop target and, via the pack's frontend, inspect the metadata of the image you drop onto it, then auto-write the detected values into the workflow's linked nodes when you queue. It reads the same A1111-style parameters and schema-tagged JSON that this pack's Image Saver writes, so the round trip is clean if you save with the pack - though it also handles images from other sources that embed similar metadata.

You control what gets applied with four independent toggles. One thing to flag up front: the toggle inputs have Chinese names in the schema (自动应用模型, 自动应用 LoRA, 自动应用推理参数, 自动应用提示词). That's intentional - the display names are the input keys so labels render even if the translation cache is stale. Translate them once and you'll never think about it again:

• Model (自动应用模型) - swaps in the checkpoint/UNet the image used, matched to a local file by name/hash.
• LoRA (自动应用 LoRA) - rebuilds the LoRA stack from the image's metadata.
• Inference parameters (自动应用推理参数) - seed, steps, sampler, cfg, size, denoise.
• Prompt (自动应用提示词) - positive and negative.

All default to on. There's also an image input in the schema - a hidden filename slot that keeps the node's widget order stable and is how the picked image is remembered; you'll interact with it via the drop/pick UI, not by typing a path.

What you need for full effect

Two supporting pieces matter. Lora Manager (willmiao/ComfyUI-Lora-Manager) must be installed for LoRA application to work - without it, that toggle has nothing to write into. And the model the image used needs to exist locally; Meta Apply can align names/hashes but can't conjure a checkpoint you don't have. Applying parameters also syncs the pack's Bypass Switch, flipping a second-pass subgraph on or off depending on whether the image's metadata carried double-sample fields - so a whole adaptive workflow snaps into the right shape from one drop.

Installing it

Ships with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/tsukino4a/ComfyUI-4A-Prompt-Manager.git
cd ComfyUI-4A-Prompt-Manager
python install.py   # or: pip install -r requirements.txt

Restart after; ComfyUI Manager users search 4A Prompt Manager. Then install Lora Manager separately if you want LoRA reuse. No models to download for any of this.

Troubleshooting

• The LoRA toggle does nothing. Almost certainly Lora Manager isn't installed, or ComfyUI wasn't restarted after you installed it.
• Labels show in Chinese. Cosmetic, not a bug - the input keys are the Chinese display names. Everything behaves normally.
• "Model not found" style behavior. The image references a checkpoint you don't have locally. Either get the model or just toggle model application off and keep the rest.
• Nothing happens at all. Remember it's an output node - it does its thing when you queue. If you're expecting an instant canvas rewrite before a run, that's not the model here.