Images are the most-tested files of all, yet generic sample sites hand you a photo with no documented properties. Here, every image records exactly what it is. Denoise, restoration, inpaint, deblur, upscale, colourization, watermark, outpaint, style-transfer, and enhance suites pair degraded inputs with clean references. Segmentation masks, visual-diff twins, EXIF/ICC ladders, multipage sprites, chart-vision, codec ladders, OCR hardcases, and soft-edge matting extend the set for modern vision QA. Clean ↔ noisy and colour ↔ greyscale pairs make before/after testing trivial.
Filter images on Browse · 452 files · 37 subcategories
The same four-object fruit still life on a cluttered background — the hard input for instance segmentation and object detection. The indexed instance mask gives per-object ground truth.
A four-object fruit still life (apple, orange, banana, grapes) on white — the easy input for instance segmentation and object detection, paired with an indexed instance mask.
An indexed (palette) instance mask for the fruit still life: background = 0 and each fruit painted with its own instance id. Per-object class and normalised bounding box are listed in the specs.
The binary ground-truth mask for the portrait subject (white = foreground) — the reference for scoring a portrait cut-out or background blur.
A cat 'portrait' on a busy background — a portrait-style subject-detection and matting input, paired with its ground-truth mask.
The binary ground-truth segmentation mask for the salient robot (white = foreground). Compute IoU / F-score of a predicted mask against this file.
A robot mascot on a cluttered background — the input for a salient-object-detection or image-matting model. Ground-truth binary mask and trimap ship alongside it.
A three-level matting trimap (black = background, grey = unknown edge band, white = foreground) for the salient robot — the standard auxiliary input for alpha-matting methods.
A 256×256 identicon-style avatar generated deterministically from the seed “ada-lovelace” — a symmetric 5×5 block pattern like the default avatars on many sites. For testing avatar rendering, placeholder generation, and image pipelines.
A 256×256 identicon-style avatar generated deterministically from the seed “alan-turing” — a symmetric 5×5 block pattern like the default avatars on many sites. For testing avatar rendering, placeholder generation, and image pipelines.
A 256×256 identicon-style avatar generated deterministically from the seed “grace-hopper” — a symmetric 5×5 block pattern like the default avatars on many sites. For testing avatar rendering, placeholder generation, and image pipelines.
A 256×256 identicon-style avatar generated deterministically from the seed “katherine-johnson” — a symmetric 5×5 block pattern like the default avatars on many sites. For testing avatar rendering, placeholder generation, and image pipelines.
A left-to-right alpha gradient over a solid colour, for testing how a viewer or compositor renders partial transparency.
The binary ground-truth alpha mask (white = subject, black = background) for the cluttered-background apple — compute IoU or boundary F-score of a predicted mask against this file.
The ground-truth transparent cut-out of the apple used in the cluttered-background hard case — the ideal output of a perfect background remover.
An apple on a busy, colour-matched background — the hard input for a background remover or matting model. Score a prediction against the paired ground-truth cut-out and alpha mask.
The expected transparent cut-out of the coffee mug, paired with the on-background version — the ground truth for a background-removal test.
A coffee mug photographed on a clean studio-white background — the easy input for a background remover, paired 1:1 with its transparent cut-out for scoring.
The coffee mug with a soft drop shadow on transparency — tests whether a background remover keeps or discards the contact shadow.
Ground-truth cut-out for the colour-matched hard case.
Subject colours deliberately match the background clutter — a hard segmentation case.
Binary mask for the colour-matched hard case.
Ground-truth RGBA cut-out for the fur-edge hard case.
Synthetic fur-like fibres on a cluttered background — a classic background-removal failure case.
Binary mask for the fur-edge hard case.
Ground-truth partial-alpha glass bottle cut-out.
Semi-transparent glass bottle on clutter — partial-alpha background-removal hard case.
Binary mask for the glass-bottle hard case.
Ground-truth mug cut-out including the transparent handle hole.
Mug with a handle interior hole on clutter — background should show through the handle.
Trimap for the handle-hole mug (black/gray/white) for matting tools.
Ground-truth mesh cut-out with transparent holes.
Chain-link / mesh with interior holes — tests whether a removers keeps background visible through the grid.
Binary mask for the mesh-fence hard case.
A potted plant with many thin leaves and stems on transparency — the canonical fine-edge test for how cleanly a background remover cuts around delicate foliage.
A half-full drinking glass rendered with partial transparency — the hard case for background removers that assume every subject pixel is opaque.
Synthetic dashboard mock (KPI cards + bar chart) — SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart) — SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart) — SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart) — SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart) — SAMPLE UI for layout / chart vision models. Not a real product.
Synthetic dashboard mock (KPI cards + bar chart) — SAMPLE UI for layout / chart vision models. Not a real product.
Compact sparkline PNG for chart-extraction / thumbnail vision tests.
Compact sparkline PNG for chart-extraction / thumbnail vision tests.
Compact sparkline PNG for chart-extraction / thumbnail vision tests.
Rendered SAMPLE data table as a PNG — for table-detection / OCR vision pipelines.
Rendered SAMPLE data table as a PNG — for table-detection / OCR vision pipelines.
Rendered SAMPLE data table as a PNG — for table-detection / OCR vision pipelines.
A clean bar chart of monthly revenue with axis labels, gridlines, and value labels — a real-world fixture for testing chart-extraction and OCR tools, vision models, and thumbnail rendering.
A two-series line chart (desktop vs. mobile visitors) with markers, a legend, and gridlines — for testing chart parsing, trend extraction, and image pipelines against a known plot.
A five-slice pie chart of traffic sources with percentage labels and a legend — a real-world fixture for chart-recognition, segmentation, and OCR of embedded text.
A scatter plot of 120 seeded points (order value versus items per order) showing a positive correlation, with labelled axes — for testing point-cloud extraction and chart parsing.
Same blocks scene encoded as png — rung on the Wave E codec ladder for quality / convert comparisons.
Same blocks scene encoded as q20 — rung on the Wave E codec ladder for quality / convert comparisons.
Same blocks scene encoded as q50 — rung on the Wave E codec ladder for quality / convert comparisons.
Same blocks scene encoded as q80 — rung on the Wave E codec ladder for quality / convert comparisons.
Same blocks scene encoded as webp — rung on the Wave E codec ladder for quality / convert comparisons.
Same fruit scene encoded as png — rung on the Wave E codec ladder for quality / convert comparisons.
Same fruit scene encoded as q20 — rung on the Wave E codec ladder for quality / convert comparisons.
Same fruit scene encoded as q50 — rung on the Wave E codec ladder for quality / convert comparisons.
Same fruit scene encoded as q80 — rung on the Wave E codec ladder for quality / convert comparisons.
Same fruit scene encoded as webp — rung on the Wave E codec ladder for quality / convert comparisons.
High-quality WebP of the fruit scene — near-lossless rung of the codec ladder.
Same mug scene encoded as png — rung on the Wave E codec ladder for quality / convert comparisons.
Same mug scene encoded as q20 — rung on the Wave E codec ladder for quality / convert comparisons.
Same mug scene encoded as q50 — rung on the Wave E codec ladder for quality / convert comparisons.
Same mug scene encoded as q80 — rung on the Wave E codec ladder for quality / convert comparisons.
Same mug scene encoded as webp — rung on the Wave E codec ladder for quality / convert comparisons.
A 16-bit (deep-colour) grayscale PNG holding a smooth 0–65535 gradient — for testing high-bit-depth support and spotting banding when a tool truncates to 8-bit.
The fruit still life as a baseline (sequential) JPEG — paired with a progressive JPEG of identical content for testing decode order and progressive-rendering support.
The fruit still life converted to the CMYK (print) colour space and saved as a JPEG — for testing CMYK decoding and CMYK→RGB conversion. Some web viewers render CMYK JPEGs with a colour shift.
A JPEG carrying rich EXIF metadata (camera make/model, software, artist, copyright, timestamp) — paired with an identical metadata-stripped copy for testing EXIF reading and privacy-scrubbing tools.
The same photo with all EXIF metadata removed — the stripped twin of the EXIF-present JPEG, for verifying that a metadata-scrubbing tool actually removed everything.
The fruit still life saved as a JPEG with an embedded sRGB ICC colour profile — paired with an untagged copy for testing colour-management and profile handling.
The pixel-identical PNG control with text and ICC chunks removed, for proving that a scrubber changes metadata without changing decoded RGB values.
A deterministic PNG with known text chunks and an embedded sRGB ICC profile, paired with a pixel-identical stripped control for extraction and privacy-tool tests.
The same photo as a progressive JPEG (loads coarse-to-fine) — the twin of the baseline JPEG, for testing progressive decoding and byte-order handling.
A losslessly compressed TIFF carrying known camera/software tags plus XMP and IPTC payloads, paired with an exactly decoded stripped image.
The same losslessly compressed TIFF pixels with EXIF-style, XMP and IPTC metadata absent, providing a ground-truth scrubber control.
The same photo saved with no embedded ICC profile — the untagged twin of the sRGB-tagged JPEG, for testing how a tool assumes/handles a missing colour profile.
True-colour ground truth for the fruit colourization suite. Compare a colourizer's output against this reference to score accuracy.
True-colour ground truth for the mug colourization suite. Compare a colourizer's output against this reference to score accuracy.
True-colour ground truth for the plant colourization suite. Compare a colourizer's output against this reference to score accuracy.
Greyscale input for colourizing the fruit scene. Pair with the colour ground truth in this group.
Sepia input for colourizing the fruit scene. Pair with the colour ground truth in this group.
Greyscale input for colourizing the mug scene. Pair with the colour ground truth in this group.
Sepia input for colourizing the mug scene. Pair with the colour ground truth in this group.
Greyscale input for colourizing the plant scene. Pair with the colour ground truth in this group.
Sepia input for colourizing the plant scene. Pair with the colour ground truth in this group.
Colour ground-truth for the apple conditioning suite. Every map in this group is derived from this exact image, so a guided model's output can be scored against it.
Colour ground-truth for the cat conditioning suite. Every map in this group is derived from this exact image, so a guided model's output can be scored against it.
Colour ground-truth for the fruit conditioning suite. Every map in this group is derived from this exact image, so a guided model's output can be scored against it.
Colour ground-truth for the mug conditioning suite. Every map in this group is derived from this exact image, so a guided model's output can be scored against it.
Colour ground-truth for the plant conditioning suite. Every map in this group is derived from this exact image, so a guided model's output can be scored against it.
Colour ground-truth for the robot conditioning suite. Every map in this group is derived from this exact image, so a guided model's output can be scored against it.
Binary Canny edge map extracted from the ground truth — feed as a structure-control input and score the generated result against the reference.
Synthetic depth map (near = bright) for depth-conditioned generation and monocular depth-estimation regression against a known reference.
Clean line-art extraction (black lines on white) for line-art-conditioned generation and line-extraction quality tests.
Tangent-space surface-normal map encoded from the scene depth — a normal-conditioning input for guided generation and relighting tests.
Sparse scribble derived from the edges — the loose hand-drawn control input for scribble-guided image generation.
Semantic segmentation map (subject vs background) — the label reference for segmentation-conditioned generation and mask-quality scoring.
Binary Canny edge map extracted from the ground truth — feed as a structure-control input and score the generated result against the reference.
Synthetic depth map (near = bright) for depth-conditioned generation and monocular depth-estimation regression against a known reference.
Clean line-art extraction (black lines on white) for line-art-conditioned generation and line-extraction quality tests.
Tangent-space surface-normal map encoded from the scene depth — a normal-conditioning input for guided generation and relighting tests.
OpenPose-style keypoint skeleton for the figure — a pose-control input for guided generation and a target for pose-estimation checks.
Sparse scribble derived from the edges — the loose hand-drawn control input for scribble-guided image generation.
Semantic segmentation map (subject vs background) — the label reference for segmentation-conditioned generation and mask-quality scoring.
Binary Canny edge map extracted from the ground truth — feed as a structure-control input and score the generated result against the reference.
Synthetic depth map (near = bright) for depth-conditioned generation and monocular depth-estimation regression against a known reference.
Clean line-art extraction (black lines on white) for line-art-conditioned generation and line-extraction quality tests.
Tangent-space surface-normal map encoded from the scene depth — a normal-conditioning input for guided generation and relighting tests.
Sparse scribble derived from the edges — the loose hand-drawn control input for scribble-guided image generation.
Semantic segmentation map (subject vs background) — the label reference for segmentation-conditioned generation and mask-quality scoring.
Binary Canny edge map extracted from the ground truth — feed as a structure-control input and score the generated result against the reference.
Synthetic depth map (near = bright) for depth-conditioned generation and monocular depth-estimation regression against a known reference.
Clean line-art extraction (black lines on white) for line-art-conditioned generation and line-extraction quality tests.
Tangent-space surface-normal map encoded from the scene depth — a normal-conditioning input for guided generation and relighting tests.
Sparse scribble derived from the edges — the loose hand-drawn control input for scribble-guided image generation.
Semantic segmentation map (subject vs background) — the label reference for segmentation-conditioned generation and mask-quality scoring.
Binary Canny edge map extracted from the ground truth — feed as a structure-control input and score the generated result against the reference.
Synthetic depth map (near = bright) for depth-conditioned generation and monocular depth-estimation regression against a known reference.
Clean line-art extraction (black lines on white) for line-art-conditioned generation and line-extraction quality tests.
Tangent-space surface-normal map encoded from the scene depth — a normal-conditioning input for guided generation and relighting tests.
Sparse scribble derived from the edges — the loose hand-drawn control input for scribble-guided image generation.
Semantic segmentation map (subject vs background) — the label reference for segmentation-conditioned generation and mask-quality scoring.
Binary Canny edge map extracted from the ground truth — feed as a structure-control input and score the generated result against the reference.
Synthetic depth map (near = bright) for depth-conditioned generation and monocular depth-estimation regression against a known reference.
Clean line-art extraction (black lines on white) for line-art-conditioned generation and line-extraction quality tests.
Tangent-space surface-normal map encoded from the scene depth — a normal-conditioning input for guided generation and relighting tests.
OpenPose-style keypoint skeleton for the figure — a pose-control input for guided generation and a target for pose-estimation checks.
Sparse scribble derived from the edges — the loose hand-drawn control input for scribble-guided image generation.
Semantic segmentation map (subject vs background) — the label reference for segmentation-conditioned generation and mask-quality scoring.
Clean reference for deblur base A. Compare deblurred outputs of the blurred variants in this group against this file.
Clean reference for deblur base B. Compare deblurred outputs of the blurred variants in this group against this file.
Base A with gauss-3 blur. Score a deblur tool against the clean reference in this group.
Base B with gauss-3 blur. Score a deblur tool against the clean reference in this group.
Base A with gauss-6 blur. Score a deblur tool against the clean reference in this group.
Base B with gauss-6 blur. Score a deblur tool against the clean reference in this group.
Base A with motion-diag blur. Score a deblur tool against the clean reference in this group.
Base B with motion-diag blur. Score a deblur tool against the clean reference in this group.
Base A with motion-h blur. Score a deblur tool against the clean reference in this group.
Base B with motion-h blur. Score a deblur tool against the clean reference in this group.
Base A with motion-v blur. Score a deblur tool against the clean reference in this group.
Base B with motion-v blur. Score a deblur tool against the clean reference in this group.
Base A with shake blur. Score a deblur tool against the clean reference in this group.
Base B with shake blur. Score a deblur tool against the clean reference in this group.
The clean reference image for denoise base A — a fruit still life on studio white. Diff the noisy variants in this set against this file to measure filter accuracy.
The clean reference image for denoise base B — a coffee mug on studio white. Diff the noisy variants in this set against this file to measure filter accuracy.
Base A with additive Gaussian noise at σ=10 (seed 4210). Compare against the clean reference in this set to score a denoise filter.
Base B with additive Gaussian noise at σ=10 (seed 7710). Compare against the clean reference in this set to score a denoise filter.
Base A with additive Gaussian noise at σ=25 (seed 4225). Compare against the clean reference in this set to score a denoise filter.
Base B with additive Gaussian noise at σ=25 (seed 7725). Compare against the clean reference in this set to score a denoise filter.
Base A with additive Gaussian noise at σ=50 (seed 4250). Compare against the clean reference in this set to score a denoise filter.
Base B with additive Gaussian noise at σ=50 (seed 7750). Compare against the clean reference in this set to score a denoise filter.
Base A with 10% salt-and-pepper noise (seed 42010). A direct fixture for testing median and morphological denoisers against the clean reference.
Base B with 10% salt-and-pepper noise (seed 77010). A direct fixture for testing median and morphological denoisers against the clean reference.
Base A with 2% salt-and-pepper noise (seed 42002). A direct fixture for testing median and morphological denoisers against the clean reference.
Base B with 2% salt-and-pepper noise (seed 77002). A direct fixture for testing median and morphological denoisers against the clean reference.
Base A with 5% salt-and-pepper noise (seed 42005). A direct fixture for testing median and morphological denoisers against the clean reference.
Base B with 5% salt-and-pepper noise (seed 77005). A direct fixture for testing median and morphological denoisers against the clean reference.
Baseline twin A for visual-diff regression (pixel change suite: 1px-crop). Compare against twin B.
Baseline twin A for visual-diff regression (pixel change suite: 1px-shift). Compare against twin B.
Baseline twin A for visual-diff regression (button change suite: button-add). Compare against twin B.
Baseline twin A for visual-diff regression (button change suite: button-label). Compare against twin B.
Baseline twin A for visual-diff regression (colour change suite: colour-tint). Compare against twin B.
Baseline twin A for visual-diff regression (button change suite: corner-badge). Compare against twin B.
Baseline twin A for visual-diff regression (pixel change suite: dot-noise). Compare against twin B.
Baseline twin A for visual-diff regression (colour change suite: hue-shift). Compare against twin B.
Changed twin B for visual-diff regression (pixel: 1px-crop). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (pixel: 1px-shift). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (button: button-add). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (button: button-label). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (colour: colour-tint). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (button: corner-badge). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (pixel: dot-noise). Expect a non-zero pixel / perceptual diff vs A.
Changed twin B for visual-diff regression (colour: hue-shift). Expect a non-zero pixel / perceptual diff vs A.
An 8-frame looping animated GIF of a moving dot — tests animation handling, first-frame extraction, and GIF parsing.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 1 (normal). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 2 (mirror horizontal). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 3 (rotate 180°). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 4 (mirror vertical). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 5 (mirror + rotate 90° CCW). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 6 (rotate 90° CW). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 7 (mirror + rotate 90° CW). Tests whether your tool honours orientation metadata.
A JPEG whose pixels are stored identically but whose EXIF Orientation tag is set to 8 (rotate 90° CCW). Tests whether your tool honours orientation metadata.
A 1-pixel-wide, 5000-pixel-tall strip — an extreme aspect ratio for stress-testing layout, scaling, and thumbnail generators.
An intentionally corrupt JPEG, truncated to half its bytes, for testing how a decoder handles incomplete image data. This is not a valid image by design.
The same 512px source saved at JPEG quality 10 — part of a q10/50/90 ladder for comparing compression artefacts at a glance.
The same 512px source saved at JPEG quality 50 — part of a q10/50/90 ladder for comparing compression artefacts at a glance.
The same 512px source saved at JPEG quality 90 — part of a q10/50/90 ladder for comparing compression artefacts at a glance.
Clean fruit still life for a second enhancement ladder subject.
Correctly exposed mug reference for scoring auto-enhance, exposure correction, and dehaze tools.
Mug photo with deliberate cool cast for testing auto-enhance and colour/exposure correction against the clean reference.
Mug photo with deliberate haze for testing auto-enhance and colour/exposure correction against the clean reference.
Hazy fruit scene for dehaze / clarity enhancement tests.
Mug photo with deliberate high contrast for testing auto-enhance and colour/exposure correction against the clean reference.
Mug photo with deliberate low contrast for testing auto-enhance and colour/exposure correction against the clean reference.
Mug photo with deliberate overexposed for testing auto-enhance and colour/exposure correction against the clean reference.
Severely underexposed fruit scene for enhancement / brighten tools.
Mug photo with deliberate underexposed for testing auto-enhance and colour/exposure correction against the clean reference.
Mug photo with deliberate warm cast for testing auto-enhance and colour/exposure correction against the clean reference.
JPEG with EXIF Orientation=2 (flip-h). Pixels are upright — viewers/tools should apply the tagged transform. SAMPLE fixture.
JPEG with EXIF Orientation=4 (flip-v). Pixels are upright — viewers/tools should apply the tagged transform. SAMPLE fixture.
JPEG with EXIF Orientation=1 (normal). Pixels are upright — viewers/tools should apply the tagged transform. SAMPLE fixture.
JPEG with EXIF Orientation=3 (rotate-180). Pixels are upright — viewers/tools should apply the tagged transform. SAMPLE fixture.
JPEG with EXIF Orientation=8 (rotate-90-ccw). Pixels are upright — viewers/tools should apply the tagged transform. SAMPLE fixture.
JPEG with EXIF Orientation=6 (rotate-90-cw). Pixels are upright — viewers/tools should apply the tagged transform. SAMPLE fixture.
JPEG with less-common EXIF Orientation=5 (transpose) for exhaustive orientation-handler coverage.
JPEG with less-common EXIF Orientation=7 (transverse) for exhaustive orientation-handler coverage.
Mug scene JPEG with an embedded sRGB ICC profile — pair with the untagged twin for colour-management tests.
PNG whose colour intent is documented as sRGB in specs (no mandatory ICC blob) — for tools that read metadata/specs.
Identical mug JPEG with no ICC profile. Specs note that sRGB is commonly assumed when untagged.
Clean synthetic face illustration (not a real person) used as ground truth for face-restore / enhancement tools.
Clean synthetic face illustration (not a real person) used as ground truth for face-restore / enhancement tools.
Clean synthetic face illustration (not a real person) used as ground truth for face-restore / enhancement tools.
Clean synthetic face illustration (not a real person) used as ground truth for face-restore / enhancement tools.
Deliberately degraded synthetic face (blur) for testing face restore / enhancement against the clean reference.
Deliberately degraded synthetic face (combo) for testing face restore / enhancement against the clean reference.
Deliberately degraded synthetic face (down) for testing face restore / enhancement against the clean reference.
Deliberately degraded synthetic face (jpeg) for testing face restore / enhancement against the clean reference.
Damage-region mask (L) for the face-restore case — white = degraded region.
Damage-region mask (L) for the face-restore case — white = degraded region.
Damage-region mask (L) for the face-restore case — white = degraded region.
Damage-region mask (L) for the face-restore case — white = degraded region.
A 1024×1024 fruit still life exported as AVIF — the modern-codec member of the conversion set.
A 1024×1024 fruit still life exported as BMP — one member of a conversion set that carries identical content across formats, so you can convert one and diff against the others.
A 1024×1024 fruit still life exported as GIF — one member of a conversion set that carries identical content across formats, so you can convert one and diff against the others.
A 1024×1024 fruit still life exported as HEIC — the HEIF/HEVC member of the conversion set, the format modern phones use for photos.
A 1024×1024 fruit still life exported as HEIF — the HEIF/HEVC member of the conversion set, the format modern phones use for photos.
A 1024×1024 fruit still life exported as JPG — one member of a conversion set that carries identical content across formats, so you can convert one and diff against the others.
A 1024×1024 fruit still life exported as PNG — one member of a conversion set that carries identical content across formats, so you can convert one and diff against the others.
A 1024×1024 fruit still life exported as TIFF — one member of a conversion set that carries identical content across formats, so you can convert one and diff against the others.
A 1024×1024 fruit still life exported as WEBP — one member of a conversion set that carries identical content across formats, so you can convert one and diff against the others.
A 16-step greyscale wedge from pure black to pure white, for testing tone reproduction, banding, and monitor calibration.
A flat 50% grey card (rgb 128,128,128), a reference for white balance and exposure testing.
A pure black-and-white checkerboard, useful for testing scaling, aliasing, and edge handling.
A continuous linear greyscale ramp from black to white, useful for spotting banding and gamma issues.
A colour fruit still life (apple, orange, banana, grapes), paired with a true-greyscale conversion — a real-world subject for testing desaturation and how distinct hues collapse to similar greys.
True 8-bit greyscale conversion of the fruit still life, the paired reference for testing colour-to-grey conversion.
Vertical bars sweeping the full hue circle at maximum saturation, paired with a greyscale version that shows how colours of different hues collapse to similar greys.
True 8-bit greyscale conversion of the full-saturation hue bars, the paired reference for testing colour-to-grey conversion.
A smooth three-channel gradient composition, paired with its greyscale conversion to test luminance handling.
True 8-bit greyscale conversion of the gradient composition, the paired reference for testing colour-to-grey conversion.
8-bit PNG twin of the linear 16-bit greyscale ramp — compare banding after truncation.
64×64 linear 16-bit checkerboard TIFF for bit-depth and downsample tests.
64×64 linear 16-bit greyscale TIFF with deflate — Wave H depth for high bit-depth loaders.
LZW twin of the Wave H linear 16-bit greyscale TIFF.
A linear 16-bit greyscale TIFF ramp (0–65535). Not a PQ HDR master — for testing high bit-depth TIFF loaders and 16→8 truncation.
A 16-bit TIFF biased into the upper code values (near clip) — for highlight-rolloff and tone-map tests on linear data.
Linear 16-bit TIFF with a dark edge and near-clip field — tone-map edge case.
Clean fruit still-life reference for the inpaint cut suite. Each cut variant and mask in this group points back to this file.
Missing-region (centre-circle) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
Missing-region (centre-rect) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
Missing-region (corner) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
Missing-region (edge-crop) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
Missing-region (irregular-tear) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
Missing-region (transparent-hole) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
Missing-region (vertical-strip) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the centre-circle cut. Feed this mask with the cut input to an inpainting tool.
White = missing region for the centre-rect cut. Feed this mask with the cut input to an inpainting tool.
White = missing region for the corner cut. Feed this mask with the cut input to an inpainting tool.
White = missing region for the edge-crop cut. Feed this mask with the cut input to an inpainting tool.
White = missing region for the irregular-tear cut. Feed this mask with the cut input to an inpainting tool.
White = missing region for the transparent-hole cut. Feed this mask with the cut input to an inpainting tool.
White = missing region for the vertical-strip cut. Feed this mask with the cut input to an inpainting tool.
Lossless PNG ground truth for JPEG blocking / ringing artefact repair tests.
Fruit still life saved at JPEG quality 10 — heavy blocking/ringing for artefact-removal and enhance tools.
Fruit still life saved at JPEG quality 15 — heavy blocking/ringing for artefact-removal and enhance tools.
Fruit still life saved at JPEG quality 25 — heavy blocking/ringing for artefact-removal and enhance tools.
Fruit still life saved at JPEG quality 40 — heavy blocking/ringing for artefact-removal and enhance tools.
Fruit still life saved at JPEG quality 5 — heavy blocking/ringing for artefact-removal and enhance tools.
Clean RGB composite of the fine-fringe soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (hair edges: fine-fringe) for alpha-matting / background-removal quality tests.
Clean RGB composite of the fur-edge soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (hair edges: fur-edge) for alpha-matting / background-removal quality tests.
Clean RGB composite of the glass-rim soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (soft edges: glass-rim) for alpha-matting / background-removal quality tests.
Clean RGB composite of the hair-strands soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (hair edges: hair-strands) for alpha-matting / background-removal quality tests.
Clean RGB composite of the plant-edge soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (soft edges: plant-edge) for alpha-matting / background-removal quality tests.
Clean RGB composite of the soft-edge-blob soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (soft edges: soft-edge-blob) for alpha-matting / background-removal quality tests.
Clean RGB composite of the wispy-top soft-alpha subject on a neutral grey — paired input for matte extractors.
Soft-alpha cut-out (hair edges: wispy-top) for alpha-matting / background-removal quality tests.
Synthetic 16×16 favicon-size mark (not a real brand) for favicon pipeline / resize tests.
Synthetic 32×32 favicon-size mark (not a real brand) for favicon pipeline / resize tests.
Synthetic 48×48 favicon-size mark (not a real brand) for favicon pipeline / resize tests.
Synthetic 64×64 favicon-size mark (not a real brand) for favicon pipeline / resize tests.
Windows ICO containing 16/32/48 favicon sizes of the synthetic mark — for ICO parsers and extractors.
Three-page TIFF of synthetic scenes — fixture for multipage TIFF loaders and page extractors.
Single-page TIFF of the same scene as page 1 of the multipage twin.
RGBA sprite sheet (2×2 cells of 64px) with numbered frames for atlas / CSS-sprite tools.
RGBA sprite sheet (3×2 cells of 64px) with numbered frames for atlas / CSS-sprite tools.
RGBA sprite sheet (4×1 cells of 64px) with numbered frames for atlas / CSS-sprite tools.
Two-page TIFF companion for multipage loader edge cases.
The vCard (.vcf) ground truth for the business-card image — the same name, title, company, email, phone, and address in structured form. The positive reference for testing OCR-to-contact extraction.
A rendered business card with a name, title, company, email, phone, and address — paired with a matching vCard so you can measure OCR and contact-extraction accuracy against a known ground truth.
Hard OCR case (dotted): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (dotted): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (low-contrast): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (low-contrast): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (low-contrast): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (skew): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (skew): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Hard OCR case (skew): SAMPLE text rendered with challenging conditions. Pair with the .txt ground truth.
Exact ground-truth string for the dotted-code OCR hard case.
Exact ground-truth string for the dotted-serial OCR hard case.
Exact ground-truth string for the low-contrast OCR hard case.
Exact ground-truth string for the low-label OCR hard case.
Exact ground-truth string for the low-receipt OCR hard case.
Exact ground-truth string for the skew-address OCR hard case.
Exact ground-truth string for the skew-date OCR hard case.
Exact ground-truth string for the skew-invoice OCR hard case.
Full-frame plant scene used as the target for outpainting tests with a 128px empty border input.
Full-frame reference for 192px border outpainting.
Full-frame plant scene used as the target for outpainting tests with a 64px empty border input.
Full-frame reference for 96px border outpainting.
Plant content centred with a 128px empty border — classic outpainting / generative expand fixture.
Content with 192px empty border for outpainting tools.
Plant content centred with a 64px empty border — classic outpainting / generative expand fixture.
Content with 96px empty border for outpainting tools.
White = region to outpaint / expand; black = known content.
Expand-region mask for outpainting.
White = region to outpaint / expand; black = known content.
Expand-region mask for outpainting.
Ambient-occlusion map darkening crevices where ambient light is blocked. Part of the brick wall PBR set — pair with the other channels in this group.
Base-colour / albedo map — the diffuse reflectance with no lighting baked in. Part of the brick wall PBR set — pair with the other channels in this group.
Height / displacement map used for parallax or true displacement of the surface. Part of the brick wall PBR set — pair with the other channels in this group.
Metallic map (black = dielectric, white = metal) selecting the reflectance model per texel. Part of the brick wall PBR set — pair with the other channels in this group.
Tangent-space normal map encoded from the surface height — perturbs shading normals for detail without extra geometry. Part of the brick wall PBR set — pair with the other channels in this group.
Roughness map (black = glossy, white = matte) driving the microfacet spread of the BRDF. Part of the brick wall PBR set — pair with the other channels in this group.
Ambient-occlusion map darkening crevices where ambient light is blocked. Part of the brushed metal PBR set — pair with the other channels in this group.
Base-colour / albedo map — the diffuse reflectance with no lighting baked in. Part of the brushed metal PBR set — pair with the other channels in this group.
Height / displacement map used for parallax or true displacement of the surface. Part of the brushed metal PBR set — pair with the other channels in this group.
Metallic map (black = dielectric, white = metal) selecting the reflectance model per texel. Part of the brushed metal PBR set — pair with the other channels in this group.
Tangent-space normal map encoded from the surface height — perturbs shading normals for detail without extra geometry. Part of the brushed metal PBR set — pair with the other channels in this group.
Roughness map (black = glossy, white = matte) driving the microfacet spread of the BRDF. Part of the brushed metal PBR set — pair with the other channels in this group.
Ambient-occlusion map darkening crevices where ambient light is blocked. Part of the ceramic tiles PBR set — pair with the other channels in this group.
Base-colour / albedo map — the diffuse reflectance with no lighting baked in. Part of the ceramic tiles PBR set — pair with the other channels in this group.
Height / displacement map used for parallax or true displacement of the surface. Part of the ceramic tiles PBR set — pair with the other channels in this group.
Metallic map (black = dielectric, white = metal) selecting the reflectance model per texel. Part of the ceramic tiles PBR set — pair with the other channels in this group.
Tangent-space normal map encoded from the surface height — perturbs shading normals for detail without extra geometry. Part of the ceramic tiles PBR set — pair with the other channels in this group.
Roughness map (black = glossy, white = matte) driving the microfacet spread of the BRDF. Part of the ceramic tiles PBR set — pair with the other channels in this group.
Ambient-occlusion map darkening crevices where ambient light is blocked. Part of the wood planks PBR set — pair with the other channels in this group.
Base-colour / albedo map — the diffuse reflectance with no lighting baked in. Part of the wood planks PBR set — pair with the other channels in this group.
Height / displacement map used for parallax or true displacement of the surface. Part of the wood planks PBR set — pair with the other channels in this group.
Metallic map (black = dielectric, white = metal) selecting the reflectance model per texel. Part of the wood planks PBR set — pair with the other channels in this group.
Tangent-space normal map encoded from the surface height — perturbs shading normals for detail without extra geometry. Part of the wood planks PBR set — pair with the other channels in this group.
Roughness map (black = glossy, white = matte) driving the microfacet spread of the BRDF. Part of the wood planks PBR set — pair with the other channels in this group.
Ambient-occlusion map darkening crevices where ambient light is blocked. Part of the woven fabric PBR set — pair with the other channels in this group.
Base-colour / albedo map — the diffuse reflectance with no lighting baked in. Part of the woven fabric PBR set — pair with the other channels in this group.
Height / displacement map used for parallax or true displacement of the surface. Part of the woven fabric PBR set — pair with the other channels in this group.
Metallic map (black = dielectric, white = metal) selecting the reflectance model per texel. Part of the woven fabric PBR set — pair with the other channels in this group.
Tangent-space normal map encoded from the surface height — perturbs shading normals for detail without extra geometry. Part of the woven fabric PBR set — pair with the other channels in this group.
Roughness map (black = glossy, white = matte) driving the microfacet spread of the BRDF. Part of the woven fabric PBR set — pair with the other channels in this group.
A 1200×630 placeholder image with its dimensions labelled on a coloured grid — a drop-in stand-in for testing layouts, responsive images, and thumbnail generation. Part of an avatar-to-Full-HD placeholder pack.
A 150×150 placeholder image with its dimensions labelled on a coloured grid — a drop-in stand-in for testing layouts, responsive images, and thumbnail generation. Part of an avatar-to-Full-HD placeholder pack.
A 1920×1080 placeholder image with its dimensions labelled on a coloured grid — a drop-in stand-in for testing layouts, responsive images, and thumbnail generation. Part of an avatar-to-Full-HD placeholder pack.
A 320×320 placeholder image with its dimensions labelled on a coloured grid — a drop-in stand-in for testing layouts, responsive images, and thumbnail generation. Part of an avatar-to-Full-HD placeholder pack.
A 640×480 placeholder image with its dimensions labelled on a coloured grid — a drop-in stand-in for testing layouts, responsive images, and thumbnail generation. Part of an avatar-to-Full-HD placeholder pack.
A 800×600 placeholder image with its dimensions labelled on a coloured grid — a drop-in stand-in for testing layouts, responsive images, and thumbnail generation. Part of an avatar-to-Full-HD placeholder pack.
A 12-frame animated PNG (APNG) of a rotating arc — a lossless, alpha-capable alternative to animated GIF. Useful for testing APNG support, frame extraction, and GIF↔APNG conversion.
A Windows ICO containing four square sizes (16/32/48/64 px) of a simple app glyph — the classic favicon/desktop-icon container. Handy for testing icon extraction and multi-size rendering.
A 256×256 fruit still-life image as binary Netpbm P4 (1-bit bitmap) — the minimal, header-plus-raw-pixels family used across Unix imaging tools. For testing Netpbm parsers and conversion.
A 256×256 fruit still-life image as binary Netpbm P5 (8-bit greyscale) — the minimal, header-plus-raw-pixels family used across Unix imaging tools. For testing Netpbm parsers and conversion.
A 256×256 fruit still-life image as binary Netpbm P6 (24-bit colour) — the minimal, header-plus-raw-pixels family used across Unix imaging tools. For testing Netpbm parsers and conversion.
A 256×256 fruit still-life image saved as uncompressed Truevision TGA — a format common in games and 3D texturing. For testing TGA decoders and conversion to modern formats.
A fruit still life rendered at 1080×1920 (vertical 9:16) with a faint registration grid — part of a ladder from 16px to 4K for testing scaling, thumbnail generation, and responsive layout on recognisable content.
A fruit still life rendered at 16×16 (16px thumbnail) with a faint registration grid — part of a ladder from 16px to 4K for testing scaling, thumbnail generation, and responsive layout on recognisable content.
A fruit still life rendered at 1920×1080 (1080p 16:9) with a faint registration grid — part of a ladder from 16px to 4K for testing scaling, thumbnail generation, and responsive layout on recognisable content.
A fruit still life rendered at 256×256 (256px) with a faint registration grid — part of a ladder from 16px to 4K for testing scaling, thumbnail generation, and responsive layout on recognisable content.
A fruit still life rendered at 2560×1440 (1440p 16:9) with a faint registration grid — part of a ladder from 16px to 4K for testing scaling, thumbnail generation, and responsive layout on recognisable content.
A fruit still life rendered at 3840×2160 (4K 16:9) with a faint registration grid — part of a ladder from 16px to 4K for testing scaling, thumbnail generation, and responsive layout on recognisable content.
Clean reference for the crease photo-restoration case. Diff a restorer's output against this file; the damaged input and mask share group img-restore-crease.
Clean reference for the dust photo-restoration case. Diff a restorer's output against this file; the damaged input and mask share group img-restore-dust.
Clean reference for the horizontal-cut photo-restoration case. Diff a restorer's output against this file; the damaged input and mask share group img-restore-horizontal-cut.
Clean reference for the scratch photo-restoration case. Diff a restorer's output against this file; the damaged input and mask share group img-restore-scratch.
Clean reference for the torn-corner photo-restoration case. Diff a restorer's output against this file; the damaged input and mask share group img-restore-torn-corner.
Clean reference for the water-stain photo-restoration case. Diff a restorer's output against this file; the damaged input and mask share group img-restore-water-stain.
Damaged input for crease restoration testing (seed 408). Pair with the clean reference and damage mask in this group.
Damaged input for dust restoration testing (seed 410). Pair with the clean reference and damage mask in this group.
Damaged input for horizontal-cut restoration testing (seed 413). Pair with the clean reference and damage mask in this group.
Damaged input for scratch restoration testing (seed 409). Pair with the clean reference and damage mask in this group.
Damaged input for torn-corner restoration testing (seed 412). Pair with the clean reference and damage mask in this group.
Damaged input for water-stain restoration testing (seed 411). Pair with the clean reference and damage mask in this group.
Black-and-white damage mask for the crease restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the dust restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the horizontal-cut restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the scratch restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the torn-corner restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the water-stain restoration case (white = damaged). Use with the damaged input and clean reference.
Class-id → label map for segmentation scene 1. Ground truth companion to the RGB and mask.
Indexed L mask for scene 1: pixel value equals class id (1=sky, 2=ground, 3=object, 4=accent).
Synthetic multi-class scene 1 (sky/ground/object/accent) for semantic-segmentation smoke tests. Paired with an indexed mask and class JSON.
Class-id → label map for segmentation scene 2. Ground truth companion to the RGB and mask.
Indexed L mask for scene 2: pixel value equals class id (1=sky, 2=ground, 3=object, 4=accent).
Synthetic multi-class scene 2 (sky/ground/object/accent) for semantic-segmentation smoke tests. Paired with an indexed mask and class JSON.
Class-id → label map for segmentation scene 3. Ground truth companion to the RGB and mask.
Indexed L mask for scene 3: pixel value equals class id (1=sky, 2=ground, 3=object, 4=accent).
Synthetic multi-class scene 3 (sky/ground/object/accent) for semantic-segmentation smoke tests. Paired with an indexed mask and class JSON.
Class-id → label map for segmentation scene 4. Ground truth companion to the RGB and mask.
Indexed L mask for scene 4: pixel value equals class id (1=sky, 2=ground, 3=object, 4=accent).
Synthetic multi-class scene 4 (sky/ground/object/accent) for semantic-segmentation smoke tests. Paired with an indexed mask and class JSON.
Class-id → label map for segmentation scene 5. Ground truth companion to the RGB and mask.
Indexed L mask for scene 5: pixel value equals class id (1=sky, 2=ground, 3=object, 4=accent).
Synthetic multi-class scene 5 (sky/ground/object/accent) for semantic-segmentation smoke tests. Paired with an indexed mask and class JSON.
Class-id → label map for segmentation scene 6. Ground truth companion to the RGB and mask.
Indexed L mask for scene 6: pixel value equals class id (1=sky, 2=ground, 3=object, 4=accent).
Synthetic multi-class scene 6 (sky/ground/object/accent) for semantic-segmentation smoke tests. Paired with an indexed mask and class JSON.
Fruit still-life content photo for neural / filter style-transfer tests. Pair with the style reference images in this group.
Style reference (comic) derived from the same scene — use as a target style when transferring the photo input.
Abstract geometric style plate for style-transfer tools that accept a separate style image.
Style reference (line-art) derived from the same scene — use as a target style when transferring the photo input.
Style reference (paint-like) derived from the same scene — use as a target style when transferring the photo input.
Style reference (posterize) derived from the same scene — use as a target style when transferring the photo input.
A 256×256 checkerboard texture that tiles seamlessly — for testing texture mapping, pattern fills, tiling, and background rendering.
A 256×256 diagonal stripes texture that tiles seamlessly — for testing texture mapping, pattern fills, tiling, and background rendering.
A 256×256 grain texture — for testing texture mapping, pattern fills, tiling, and background rendering.
A 256×256 grid texture that tiles seamlessly — for testing texture mapping, pattern fills, tiling, and background rendering.
A 256×256 polka dots texture that tiles seamlessly — for testing texture mapping, pattern fills, tiling, and background rendering.
Bicubic downscale to 128×128 from the 1024px reference. Use as a super-resolution input.
256×256 bicubic downscale further JPEG-compressed at quality 35 — a harder super-resolution input with artefacts.
Bicubic downscale to 256×256 from the 1024px reference. Use as a super-resolution input.
Bicubic downscale to 64×64 from the 1024px reference. Use as a super-resolution input.
Bicubic downscale to 128×128 from the 1024px reference. Use as a super-resolution input.
256×256 bicubic downscale further JPEG-compressed at quality 35 — a harder super-resolution input with artefacts.
Bicubic downscale to 256×256 from the 1024px reference. Use as a super-resolution input.
Bicubic downscale to 64×64 from the 1024px reference. Use as a super-resolution input.
Clean 1024×1024 reference for super-resolution base A. Upscale the smaller inputs in this group toward this ground truth.
Clean 1024×1024 reference for super-resolution base B. Upscale the smaller inputs in this group toward this ground truth.
An animated loading spinner using SMIL (animateTransform) — a self-contained animated SVG for testing whether a renderer or converter handles SVG animation, and how it rasterises an animated frame.
A bar chart drawn as vector SVG — axis, gridlines, labelled bars — the scalable counterpart to the raster chart images, for testing SVG chart rendering and vector conversion.
A flat vector landscape — gradient sky, sun, layered hills, and trees built from bezier paths — a richer SVG for testing gradient and path rendering, thumbnailing, and vector conversion.
A generic sample brand logo as scalable vector graphics — a gradient mark plus a wordmark — for testing SVG rendering, rasterisation, and vector-to-raster conversion. Opens in the in-browser editor.
A single stroke-style UI icon (a document) as a 24×24 SVG — the kind of vector icon used in interfaces, for testing icon rendering, recolouring, and SVG-to-PNG conversion.
Clean fruit still life without marks — ground truth for synthetic SAMPLE watermark removal tests.
Clean fruit still life without marks — ground truth for synthetic SAMPLE watermark removal tests.
Clean fruit still life without marks — ground truth for synthetic SAMPLE watermark removal tests.
Clean fruit still life without marks — ground truth for synthetic SAMPLE watermark removal tests.
Clean fruit still life without marks — ground truth for synthetic SAMPLE watermark removal tests.
Fruit still life with a synthetic SAMPLE watermark (not a real brand). Use with the clean reference and mask to score removers.
Fruit still life with a synthetic SAMPLE watermark (not a real brand). Use with the clean reference and mask to score removers.
Fruit still life with a synthetic SAMPLE watermark (not a real brand). Use with the clean reference and mask to score removers.
Fruit still life with a synthetic SAMPLE watermark (not a real brand). Use with the clean reference and mask to score removers.
Fruit still life with a synthetic SAMPLE watermark (not a real brand). Use with the clean reference and mask to score removers.
Mask of the synthetic SAMPLE watermark region for supervised removal tests.
Mask of the synthetic SAMPLE watermark region for supervised removal tests.
Mask of the synthetic SAMPLE watermark region for supervised removal tests.
Mask of the synthetic SAMPLE watermark region for supervised removal tests.
Mask of the synthetic SAMPLE watermark region for supervised removal tests.
Every image ships with documented specs (dimensions, seeds, damage type, pair/group IDs). Denoise, restoration, colourization, watermark, and enhance suites include clean ground-truth twins so you can score tools with PSNR/SSIM instead of guessing.
No. Face-restore fixtures use synthetic illustrations, and watermark suites use SAMPLE logos/text only. Corrupt samples are labelled intentionally corrupt.
Open any suite member and use the Pair / Group links on the file page, or filter Browse by purpose (for example colourization-testing or watermark-removal-testing).
Wave E adds segmentation scene/mask/JSON triples, A/B diff-regression twins, EXIF Orientation + ICC colour notes, multipage TIFF / sprite / favicon packs, chart-as-image mocks, codec quality ladders, hard OCR cases, and soft-alpha matting hardcases.
We use Google Analytics and show ads via Adsterra. Non-essential cookies and ad scripts run only after you allow the matching categories. See our cookie policy.
