This dataset contains algorithmically generated image pairs designed to teach diffusion-based image editing models pixel-level identity preservation — the ability to keep unchanged regions of an image exactly intact while applying targeted edits.

Every example consists of a reference image, a target image, and a short natural-language prompt. The transformation between reference and target is always a deterministic, rule-based operation (no generative model was used at any stage). This guarantees perfect ground truth with zero hallucination and makes the dataset fully reproducible from any source image collection.

The dataset is intended as a complementary augmentation layer alongside semantic edit datasets such as molbal/multi_reference_image_editing. On its own it teaches preservation and spatial reasoning; combined with semantic edit data it closes the gap between "follows the instruction" and "keeps everything else pixel-perfect."

Motivation

Models trained exclusively on semantic edit pairs learn to always apply a transformation. They have never seen a training signal that says "the reference pixels must survive unchanged in the output." The consequences are visible at inference time as:

Face identity drift even on minor edits
Background detail loss
Texture and colour shifts outside the edited region

This dataset provides three complementary curriculum signals to fix that:

Signal	What it teaches
Identity copy	Full preservation — output must equal reference
Spatial transforms	Partial preservation with geometric change
Photometric transforms	Preservation under tone/colour/degradation

Source Images

All source images are the 10,000 real-world photographs from molbal/multi_reference_image_editing, which were originally sourced from Pexels with the longest edge capped at 2048 px. No additional imagery was introduced.

Pair Types

1. Identity Copy

Field	Value
Prompt	(empty string)
Reference	Original image
Target	Identical copy of reference

The prompt is intentionally empty. This forces the model to learn that the reference latent itself carries the preservation signal rather than any text instruction. These examples are the strongest anchor for pixel-level fidelity.

2. Spatial Transforms

All spatial pairs use the following convention for operations that would reduce sharpness through interpolation:

The target is always the sharp image. The reference is the geometrically transformed (potentially softer) version. This means the model always has to produce a sharp output and never has to hallucinate detail it was not given.

2a. Pan / Shift

Two overlapping crops from the same image, simulating a camera pan. The overlapping region (~65–80% of the frame) must be reproduced exactly; the non-overlapping region must be plausibly completed.

Slug	Prompt
`pan-right`	pan the view to the right
`pan-left`	pan the view to the left
`pan-down`	pan the view downward
`pan-up`	pan the view upward
`pan-right-s`	shift the view slightly to the right
`pan-left-s`	shift the view slightly to the left
`pan-down-s`	shift the view slightly downward
`pan-up-s`	shift the view slightly upward

Shift fractions: 30% (full pan) and 20% (slight shift). Output dimensions are derived from the crop aspect ratio and snapped to the nearest 64-pixel boundary. No rotation or distortion is introduced.

2b. Directional Zoom-In

Reference = full image resized to output dimensions (slightly soft). Target = directional crop resized to output dimensions (sharp).

Slug	Prompt
`zoom-in-center-1x5`	zoom in on the centre
`zoom-in-center-2x`	zoom in close on the centre
`zoom-in-left`	zoom in on the left side
`zoom-in-right`	zoom in on the right side
`zoom-in-top`	zoom in on the top part
`zoom-in-bottom`	zoom in on the bottom part

Crop factors: 60% of the relevant axis for directional zooms, 67% and 50% for centre zooms.

2c. Zoom-Out

Reference = centre crop (zoomed-in view the model sees). Target = full image (always sharp, what the model must produce).

Slug	Prompt
`zoom-out-1x5`	zoom out to reveal more of the scene
`zoom-out-2x`	zoom out to show the full scene

2d. Flip

Slug	Prompt
`flip-h`	mirror the image horizontally
`flip-v`	flip the image vertically

3. Photometric Transforms

Reference = original image. Target = transformed image. (Exception: JPEG pairs — see below.)

3a. Brightness

Slug	Prompt	Factor
`brightness-darker`	darken the image	0.55
`brightness-brighter`	brighten the image	1.45

3b. Contrast

Slug	Prompt	Factor
`contrast-low`	reduce the contrast	0.45
`contrast-high`	increase the contrast	1.85

3c. Saturation

Slug	Prompt	Factor
`saturation-bw`	convert to black and white	0.0
`saturation-desat`	desaturate the image	0.35
`saturation-vivid`	make the colors more vivid	1.85

3d. Colour Temperature

Slug	Prompt
`temp-warm`	make the image warmer
`temp-cool`	make the image cooler

Warm: red channel ×1.20, blue ×0.80. Cool: inverse.

3e. Blur

Slug	Prompt	Radius
`blur-slight`	add a slight blur	2
`blur-medium`	blur the image	5
`blur-heavy`	heavily blur the image	10

3f. Tilt-Shift

A horizontal band (30–70% of image height) is kept sharp; top and bottom are Gaussian blurred (radius 8) with 30-pixel feathering.

Slug	Prompt
`tiltshift`	apply tilt-shift effect

3g. Sharpen

Unsharp mask (radius 2, strength 160%, threshold 3).

Slug	Prompt
`sharpen`	sharpen the image

3h. Noise

Additive Gaussian noise.

Slug	Prompt	Std
`noise-slight`	add slight noise to the image	12
`noise-medium`	add noise to the image	35
`noise-heavy`	add heavy grain to the image	65

3i. JPEG Compression / Restoration

Inverted convention: reference = JPEG-compressed image, target = original sharp image. Prompt instructs removal of artifacts.

Slug	Prompt	Quality
`jpeg-heavy`	remove JPEG compression artifacts	15
`jpeg-severe`	remove severe JPEG compression artifacts	5

3j. Vignette

Radial darkening mask applied multiplicatively.

Slug	Prompt	Strength
`vignette-medium`	add vignette	0.55
`vignette-heavy`	add strong vignette	0.80

3k. Colour Tint Overlay

RGBA overlay composited onto the image.

Slug	Prompt	Colour (RGBA)
`tint-warm-orange`	add a warm orange tint	(255, 200, 150, 75)
`tint-cool-blue`	add a cool blue tint	(150, 200, 255, 75)
`tint-green`	add a green tint	(200, 255, 200, 60)
`tint-dark`	add a dark overlay	(0, 0, 0, 110)

3l. Film Grain

Multiplicative noise sampled from uniform(0.82, 1.18) per channel per pixel.

Slug	Prompt
`filmgrain`	add film grain

3m. Pixelate

Nearest-neighbour downsample then upsample.

Slug	Prompt	Factor
`pixelate-medium`	pixelate the image	8
`pixelate-heavy`	heavily pixelate the image	16

File Structure

All files sit in a single flat directory. No subdirectories.

Downloads last month: -

Total file size:

76.5 GB