Guyanonymous’ Image and Prompt Comparison Studio v6.0
A single-file, fully offline studio for comparing AI image generations and finding out which settings and wildcards actually work.
I built this for my own ComfyUI / Forge Neo workflow and it grew into something worth sharing. It’s one HTML file. Open it in Chrome or Edge, point it at your output folders, and compare, rate and sort your generations. No install, no server, no account, and it makes no network requests, so nothing is ever uploaded.
Why
Batch generation produces hundreds of images, and it’s hard to tell what made the good ones good. Was it the checkpoint, the CFG, or the wildcard that picked “linocut print”? Did the model even render the glasses, the outfit or the setting you asked for?
This lines the images up, reads the settings saved inside each one, works out which part of each prompt changed, and keeps score as you rate.
What it does
Comparing
Independent panels: 2–4 folders side by side, each from its own starting image, stepping together with the arrow keys, mouse wheel or mouse back/forward buttons
Grouped set: one big batch split into panels by a fixed count, or by any settings you choose: checkpoint, CFG, a single ComfyUI node input, or a wildcard in the prompt
Review mode: 1, 2, 4, 6 or 9 images at a time. Rate with the 1–5 keys and it moves on to the next. Order by unrated first, rating, or shuffled
Synced zoom and pan, a wipe slider, flicker, and a difference heatmap between two panels
Loupe that magnifies the same spot in every panel. Place each one separately (on each face’s eyes, say), then lock them together and move them as a group
Settings and prompts
Reads embedded settings from ComfyUI, A1111, Forge / Forge Neo, InvokeAI and NovelAI (PNG, JPEG, WebP)
Panels label themselves by what differs (“CFG 7, juggernautXL”), and the settings that differ are highlighted
Prompt diff: long wildcard prompts are sorted into topics (Subject, Eyes, Outfit, Eyewear, Pose, Setting, Lighting, Lens, Art style and about 30 more), even when your wildcards word the same thing differently. Shows what’s the same in every prompt, every value that repeats, and the panels side by side. Simple one-line templates get numbered parts instead
Align by seed or filename number, so a skipped image doesn’t throw the panels out of step
Deciding
“Did it render?” checklist: every part of the image’s prompt, marked rendered or not rendered, plus rating, flag and notes. One click marks the rest as rendered, so you only click the misses
Results by value: the average rating and render rate for every wildcard value and every setting. Click a row to review exactly those images
Filters: text in the whole prompt, in one topic or in one setting; rating; flagged; images with a part that didn’t render
Session file: ratings, notes and ticks auto-save to a file next to your images, and reconnect next time
Exporting
Grid export at up to full resolution: XY-plot style, one row per step, one grid per panel, or one file per step. PNG with no browser size cap, or JPEG / WebP (tiled when over their limits)
A labeled comparison image of the current step, and a ratings CSV with notes and render counts
Comfort
Thumbnail strip, contact sheet, autoplay, synced video (MP4 / WebM), drag and drop, and showing one panel full size by holding its number
Remembered folders and session, Watch folders (new images appear while a batch is still generating), and file pickers that reopen where you left off
How to start
Unzip and open the .html file in Chrome or Edge
Pick a mode:
Review: Select folder…, rate with 1–5, press K to mark what rendered, then press R for results
Independent panels: Choose folder… on each panel, pick the starting images, then step with the arrow keys
Grouped set: Select files…, choose By settings change, then click only next to the setting to compare
Press ? for every shortcut
The included README covers everything in about five minutes.
Version history
v6.0: Review mode; prompt diff that sorts wildcard sentences into topics; “did it render?” checklist; results by value; filters; auto-saving session file; group by prompt topic; CSV with notes and render counts; ratings matched by name and size
v5.1: Renamed Guyanonymous’ Image and Prompt Comparison Studio, with a version number; run details (time taken, hardware, hashes) no longer treated as settings
v5.0: Grid export at up to full resolution: four layouts; PNG beyond the browser’s canvas limit; JPEG / WebP tiling; labels, colours and progress bar
v4.1: Loupe with crosshair; separate and locked-together positions; size and magnification bar
v4.0: Prompt diff for wildcard templates; group by prompt part; pickers remember the last folder
v3.1: Group by any single setting or ComfyUI node input; mouse wheel and back/forward buttons step through images
v3.0: Group by chooser showing how often each setting changes
v2.0: Settings reading and labels; compare (wipe / flicker / difference); thumbnails and contact sheet; ratings and CSV; remembered sessions; auto-grouping; align by seed; watch folders; loupe; solo panel; export image; drag and drop; video
v1.0: First version: 2–4 independent panels or a grouped set, all stepping together
Good to know
Settings come from the metadata your generator saves. Some upload sites and image editors strip it, so use your original output files
Topic sorting works from the wording of English prompts. Sentences it can’t place go under their opening words or “Other text”
Ratings, notes and ticks are kept in your browser and matched by file name and size. Use the session file to keep them safe or move them to another computer
Use Chrome or Edge for remembered folders, folder watching, session auto-save and full-size PNG grids. Other browsers handle everything else
Free to use and share. Feedback, bug reports and feature ideas welcome.