Target text
Blue outlines show the target. Orange outlines show your observed location.
Layout experiment 02 / Blind A/B
Which image
followed the brief?
Give this page one target layout and two images you generated from it. It hides the Base and SFT labels while you transcribe visible text and mark where it landed.
Manual evaluation in your browser. Images stay on your device. The exported result contains image hashes and annotations, never image bytes.
01 / Prepare a pair
One target. Two outputs.
Use matching generation dimensions and settings for the two images. Choose the Base and SFT outputs separately; their display order is randomized after loading.
1. Import the layout
Paste an editable Studio draft or a compiled MOSS layout prompt. A compiled prompt omits dimensions, so enter the original generation size below.
The existing Studio parser checks field names, coordinates and exact text. Open the layout editor ↗
2. Add the outputs
Choose one Base output and one SFT output in PNG, JPEG or WebP format. File names are hidden during annotation.
No images selected
Each image must match the layout's original width and height. Cropped or resized files need their original full-size versions for this check.
02 / Observe without labels
Mark what you can see
Draw a rectangle around each visible text region, then transcribe its exact characters. Use the numeric box fields for keyboard entry. Mark missing or illegible text when needed.
Text region 1 of 10 of 2 observations complete
AIdentity hidden
Drag across the image to mark the observed text, or enter its box above.
BIdentity hidden
Drag across the image to mark the observed text, or enter its box above.
03 / This pair
What the annotations show
These figures describe two images and human observations. They are not OCR output or a general Base-versus-SFT benchmark.
Exact text keeps Unicode and line breaks. Missing text counts as a full miss. Illegible text is reported separately and excluded from averages. Placement compares your drawn box with the requested box.
The JSON includes source revisions, the validated layout, image SHA-256 digests, manual annotations, settings and randomized A/B mapping. Image pixels remain local.
Pinned sources