PROMPT Workflow 作成者: Thinking Bells

PROMPT Workflow inspects the images you come across while browsing and reveals
the AI generation data hidden inside them.

Most AI image tools write their settings into the file itself. Your browser
never shows you any of it. This extension reads those bytes directly, tells you
what kind of data is in there, and — when there is something usable — turns it
into a PNG you can drop straight onto the ComfyUI canvas.

HOW YOU USE IT

Click the toolbar button and a small WF badge appears on every image on the
page. Click a badge and a panel opens with the verdict, the raw fields, and a
single button to save the result. You can also right-click any image and choose
PROMPT Workflow from the context menu.

WHAT IT READS

• PNG text chunks: tEXt, zTXt and iTXt, including the compressed variants that
  most metadata viewers skip
• JPEG: EXIF tags and the JPEG comment (COM) marker
• WebP: the EXIF chunk

UserComment fields written by A1111 use a UTF-16BE encoding header that many
parsers get wrong, producing garbled text. PROMPT Workflow decodes them
correctly.

WHAT IT TELLS YOU

Every image gets one clear, colour-coded verdict:

• ComfyUI workflow — a full node graph, or the API-only format
• A1111 / Forge parameters — with the exact variant detected from the version
  string, so you know whether it came from Automatic1111, Forge or reForge
• SwarmUI parameters
• Metadata only — camera EXIF, editing software, or another generator such as
  InvokeAI, Fooocus, NovelAI or Midjourney
• No metadata at all

For A1111 and Forge images there is a readable view that separates the positive
prompt, the negative prompt, and a table of every parameter. The parser handles
nested braces correctly, so values like Hashes: {"model": "abc"} stay intact
instead of being split into fragments.

WHAT IT CONVERTS

One button, four different behaviours depending on what was found:

• A ComfyUI PNG is copied byte for byte, completely untouched
• A ComfyUI JPEG or WebP is repackaged into a PNG carrying the original graph,
  because ComfyUI only accepts PNG and WebP when you drag a file onto it
• The API-only format is rebuilt into a visual node graph, with the layout
  calculated by dependency depth
• A1111 and Forge parameters become a txt2img workflow: samplers and schedulers
  are mapped to their ComfyUI equivalents, and LoRAs written as <lora:name:0.8>
  in the prompt are extracted into chained LoraLoader nodes
• SwarmUI parameters become a workflow too, including Kohya Deep Shrink when
  the image used it

The saved file keeps the original filename. URLs from forums and galleries that
mangle the name — turning photo.png into photo-png.947917 inside a query string
— are resolved back to the real filename.

HONEST ABOUT WHAT IT CANNOT DO

Converting from A1111 or SwarmUI does not recover an original workflow, because
none was ever stored. It builds a new one from the parameters. The result will
look similar to the source image, not identical: ComfyUI and A1111 implement
CFG, token weighting and CLIP skip differently.

The extension lists exactly what could not be carried over — Hires. fix,
ControlNet, ADetailer, refiners, Skimmed CFG, variation seeds, batch counts —
instead of silently dropping them and letting you wonder why the output looks
wrong.

It is equally direct when there is nothing to find. Many sites re-encode
uploaded images and strip their metadata, so a lot of pictures simply carry no
generation data. When that happens the extension says so and explains the
likely reason, rather than guessing.

PRIVACY

No data is collected or transmitted anywhere. The only network request is for
the image you clicked on. There is no analytics, no telemetry, no update server
and no third-party endpoint.

The source is plain JavaScript: no dependencies, no bundler, no minification.
What you download is what runs.