MetaMan Extract Components

The middle child of the MetaMan pack, and honestly the one you'll reach for most. MetaManLoadImage hands you a big JSON blob full of someone's generation settings. MetaManExtractComponents takes that blob and cracks it open into individual typed outputs - the prompt, the seed, the steps, the CFG - so you can wire a stranger's recipe straight into your own KSampler.

This is the node that makes "steal the settings from any image" a five-minute job instead of a squinting contest over a parameter string.

What it's for

Every workflow starts the same way: you want to reproduce or riff on something. In ComfyUI's golden era, sharing an image meant sharing the pipeline - the whole graph lives in the PNG metadata (the "workflow included" culture is real). This node is how you go the other direction: image in, individual settings out, ready to plug into CLIPTextEncode, KSampler, and EmptyLatentImage.

How it works

It's a pure parser - no image processing at all. Feed it a JSON string and it reads the fields, fills in sensible defaults where something's missing, and returns fourteen values. Two details are worth knowing:

• De-obfuscation. Tensor.AI renames models to junk strings when it exports. The node prefers a model_name_real field over the raw model_name when it's present, which is a genuinely thoughtful touch for a pack this size.
• LoRA simplification. LoRAs come out as a list of {real_name, weight} pairs - the info you actually need, without the hash noise.

If the metadata is empty or unparseable it doesn't error; it hands you defaults: 20 steps, 7.0 CFG, euler, seed -1, 512×512. Fine when a setting was never recorded, slightly dangerous when you forget it did that and wonder why everything regenerates as a 512² blur.

Inputs and outputs that matter

One input: metadata_json (STRING). It's forceInput - there's no text box, you must wire it. Feed it from MetaManLoadImage's string output or any other JSON source.

The outputs, in the order that matters:

• positive_prompt and negative_prompt → the two CLIPTextEncode nodes.
• steps, cfg_scale, seed, sampler, scheduler, denoising_strength → KSampler (and KSamplerAdvanced if you're doing two-pass work).
• width, height → EmptyLatentImage.
• model_name, vae_name → informational; you still pick the actual checkpoint yourself.
• loras, embeddings → wildcard (*) types.

The honest caveats

• Sampler names are A1111-isms. The sampler output comes back like DPM++ 2M Karras, which is not a valid value in KSampler's dropdown - ComfyUI wants dpmpp_2m + karras. Treat the extracted strings as a hint, not a wire-in-ready value, and pick the matching entries yourself.
• The wildcard outputs are mostly for reading. loras and embeddings won't connect cleanly to a LoraLoader (which wants a concrete model-name string). Read them, load the files manually, move on.
• It doesn't load models. model_name is just a string; you still drag the checkpoint into your loader. The "recreate this image" promise is you, wiring these values up.

Installing and wiring

Same story as the rest of the pack: ComfyUI Manager (search "MetaMan") or

cd ComfyUI/custom_nodes
git clone https://github.com/MatthewClayHarrison/ComfyUI-MetaMan

then restart. No model downloads, no heavy dependencies - the parser is pure Python on top of what ComfyUI already ships. One heads-up: the README describes a "MetaMan Universal" mega-node that never made it into the code. What's real is this node, MetaManLoadImage, and MetaManEmbedAndSave. Use those and you're on solid ground.