Find files, editable templates and browser test targets by what you need to make or test. The directory below is cut by format; the two collections under it cut the same library by subject and by workflow.
A dense pose map for the published plate sp-climbing.png, 512x512. DensePose body-part segmentation, which labels regions of every person it finds rather than returning a skeleton, so background figures are labelled as well as the subject. Lit coverage measures 100.0% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A depth map for the published plate sp-climbing.png, 1024x1024. Monocular depth from Depth Anything V2 (vitl), rendered at the plate's long edge rather than the 512-pixel default, so depth and colour can be compared per pixel without a resample in between. Lit coverage measures 81.0% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A tangent-space surface normal map for the published plate sp-climbing.png, 1024x1024 - the same size as the plate, so the two compare pixel for pixel with no resample in between. RGB encodes the XYZ surface direction remapped from -1..1 into 0..255. Decoded back to vectors this file measures a mean length of 0.9974, which is the number to check your own decode against: a reader that transposes the channels or inverts the remap still produces a plausible-looking image, and lights the surface the wrong way.
A canny edge map for the published plate sp-cyclists.png, 1024x1024. Canny edge detection at low threshold 0.1 and high 0.3, run at the plate's own resolution so the edges land on the same pixels as the photograph they came from. Lit coverage measures 11.5% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A clean 512x512 crop of group of road cyclists on a country road, daylight, taken from the published plate sp-cyclists.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
A dense pose map for the published plate sp-cyclists.png, 512x512. DensePose body-part segmentation, which labels regions of every person it finds rather than returning a skeleton, so background figures are labelled as well as the subject. Lit coverage measures 100.0% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A depth map for the published plate sp-cyclists.png, 1024x1024. Monocular depth from Depth Anything V2 (vitl), rendered at the plate's long edge rather than the 512-pixel default, so depth and colour can be compared per pixel without a resample in between. Lit coverage measures 93.3% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
The same view of group of road cyclists on a country road, daylight, with a foreign object composited over 9.72% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
Group of road cyclists on a country road, daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 44.39/255 and from the ground-truth plate by 13.229/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 6.53/255, which is the full-frame VAE round trip and not an edit.
The exact footprint of the object sitting over group of road cyclists on a country road, daylight, as an 8-bit mask covering 9.72% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
A tangent-space surface normal map for the published plate sp-cyclists.png, 1024x1024 - the same size as the plate, so the two compare pixel for pixel with no resample in between. RGB encodes the XYZ surface direction remapped from -1..1 into 0..255. Decoded back to vectors this file measures a mean length of 0.996, which is the number to check your own decode against: a reader that transposes the channels or inverts the remap still produces a plausible-looking image, and lights the surface the wrong way.
A canny edge map for the published plate sp-football.png, 1024x1024. Canny edge detection at low threshold 0.1 and high 0.3, run at the plate's own resolution so the edges land on the same pixels as the photograph they came from. Lit coverage measures 3.8% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A clean 512x512 crop of amateur football match on a grass pitch, players contesting the ball, daylight, taken from the published plate sp-football.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
A dense pose map for the published plate sp-football.png, 512x512. DensePose body-part segmentation, which labels regions of every person it finds rather than returning a skeleton, so background figures are labelled as well as the subject. Lit coverage measures 100.0% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A depth map for the published plate sp-football.png, 1024x1024. Monocular depth from Depth Anything V2 (vitl), rendered at the plate's long edge rather than the 512-pixel default, so depth and colour can be compared per pixel without a resample in between. Lit coverage measures 85.2% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
The same view of amateur football match on a grass pitch, players contesting the ball, daylight, with a foreign object composited over 16.40% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
Amateur football match on a grass pitch, players contesting the ball, daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 50.428/255 and from the ground-truth plate by 59.266/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.641/255, which is the full-frame VAE round trip and not an edit.
The exact footprint of the object sitting over amateur football match on a grass pitch, players contesting the ball, daylight, as an 8-bit mask covering 16.40% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
A tangent-space surface normal map for the published plate sp-football.png, 1024x1024 - the same size as the plate, so the two compare pixel for pixel with no resample in between. RGB encodes the XYZ surface direction remapped from -1..1 into 0..255. Decoded back to vectors this file measures a mean length of 0.9961, which is the number to check your own decode against: a reader that transposes the channels or inverts the remap still produces a plausible-looking image, and lights the surface the wrong way.
A canny edge map for the published plate sp-gym.png, 1024x1024. Canny edge detection at low threshold 0.1 and high 0.3, run at the plate's own resolution so the edges land on the same pixels as the photograph they came from. Lit coverage measures 3.1% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
A depth map for the published plate sp-gym.png, 1024x1024. Monocular depth from Depth Anything V2 (vitl), rendered at the plate's long edge rather than the 512-pixel default, so depth and colour can be compared per pixel without a resample in between. Lit coverage measures 88.7% of the frame. The plate and this map are the same scene at the same size, so they can be compared pixel for pixel rather than by eye.
The same view of person lifting weights in a gym, indoor lighting, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 2.509/255, the VAE round trip, but over the whole rectangle by 4.2/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of person lifting weights in a gym, indoor lighting - the frame an outpainting tool is given, cropped from the published plate nss-sp-gym_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
A tangent-space surface normal map for the published plate sp-gym.png, 1024x1024 - the same size as the plate, so the two compare pixel for pixel with no resample in between. RGB encodes the XYZ surface direction remapped from -1..1 into 0..255. Decoded back to vectors this file measures a mean length of 0.9973, which is the number to check your own decode against: a reader that transposes the channels or inverts the remap still produces a plausible-looking image, and lights the surface the wrong way.
