Model-Prompt from Metadata

Here's the flagship of the Model and Prompt from Metadata pack, and the one that makes "reproduce this image" stop being an archaeological dig. Model-Prompt from Metadata (ImageMetadataPromptLoader) reads one image - or a workflow JSON - and pulls out the checkpoint, the VAE, and the positive and negative prompts in one go. It even encodes the prompts for you. Drop the PNG, everything's wired.

The day I got this working I deleted a genuinely awful habit of keeping a "master" workflow with text files taped to it. This is the "workflow included" culture - ComfyUI embeds the full generation graph in the PNG's prompt chunk - turned into a loader you don't have to think about.

How it works

Same front-end engine as its siblings: the browser-side JS parses the metadata (PNG prompt chunk, WebP EXIF, or a pasted JSON workflow), fuzzy-matches the detected checkpoint against your installed files, and shows ✓ / ✗ badges. Where this node goes further is the prompts. If exactly one prompt is detected, it's auto-selected; if the workflow used several, you click the one you want and the full text appears below for a sanity check. Everything flows into the node's inputs, and on the Python side the text is encoded with the loaded model's own CLIP - clip.tokenize + encode_from_tokens under the hood. That's the detail that makes it a faithful reproduction rather than an approximation: the conditioning matches what the model actually saw.

Inputs and outputs that matter

The only inputs you'd normally touch:

• ckpt_name and vae_name - auto-selected when exactly one match is found; None VAE means use the checkpoint's built-in one.
• positive_text / negative_text - multiline fields that fill themselves when you click a detected prompt. Editable after the fact, which is the whole point.

The _metadata_json and _image_file fields are internal plumbing. Now the good part - the outputs:

• model, clip, vae - straight to your sampler.
• positive / negative - CONDITIONING, already encoded. This node replaces both your checkpoint loader and your two CLIPTextEncode nodes.
• positive_text / negative_text - the raw strings, so you can feed them into a text-editing node (the pack's CLIP Text Encode edit+ is built for exactly this).
• metadata and image - the parsed JSON and the dropped image as an IMAGE tensor, for reference workflows.

A full txt2img workflow is basically: this node → KSampler → VAE Decode. Three nodes, done.

Installing it

Shared with the rest of the pack: ComfyUI Manager (search "Model and Prompt from Metadata") or

cd ComfyUI/custom_nodes
git clone https://github.com/ketle-man/model-and-prompt-from-metadata

then restart. The pack declares zero Python dependencies - it only uses what ComfyUI already bundles - and needs no model downloads. Lightest install in the ecosystem, honestly.

Gotchas

The README is upfront that UNet-based models (Flux, Qwen, Z-Image) are out of scope - the node will show you the detected names but can't load them; that's what the author's Workflow Studio pack is for. Second, an auto-select only happens when there's exactly one detected prompt; multi-prompt workflows need one click. And the classic trap: if the checkpoint that made your image isn't installed, it shows ✗ and nothing auto-selects - you'll pick a stand-in from the dropdown and the prompts will encode against that model's CLIP. For SD1.5/SDXL/Illustrious images you own, this node is a genuine time-saver.