Datasets:

stemjara
/

minecraft-vanilla-frames

license: other
license_name: mixed
license_details: >-
  Annotations CC BY 4.0; images are Minecraft screenshots under Mojang's usage
  guidelines; Sundial Lite renders made with a GPL-3.0 shader
pretty_name: >-
  Minecraft vanilla frames with depth, camera, traced sun shadows and Sundial
  Lite renders
task_categories:
  - image-to-image
  - depth-estimation
tags:
  - minecraft
  - webdataset
  - depth
  - shadows
  - shaders
size_categories:
  - 10K<n<100K
configs:
  - config_name: preview
    default: true
    data_files:
      - split: train
        path: preview/train-*.tar
      - split: test
        path: preview/test-*.tar

Minecraft vanilla frames with depth, camera, traced sun shadows and Sundial Lite renders

20 787 frames of Minecraft Java Edition 1.21.1 rendered without shaders at 1280×720, each with its depth buffer, its camera, the block under every pixel and the sun shadow traced in the saved world, and for 16 363 of them the same frame rendered with the Sundial Lite shader pack. They were captured with a Fabric mod that teleports the player through a plan of poses with the world frozen (no weather, fixed time of day).

Capture tool (the mod, its configuration, the shader options and the contribution validator): https://github.com/stevenijm777/nrmc-capture

Resumen en español: 20 787 fotogramas de Minecraft 1.21.1 sin shaders, con profundidad, cámara, bloque por píxel y la sombra del sol trazada en el mundo guardado; 16 363 de ellos también con el shader Sundial Lite, para aprender el paso de vanilla a shader.

Paired renders: Sundial Lite

sundial0/ holds the target half for the first capture round: the same pose, time and camera rendered with Sundial Lite v1.2.0 (Modrinth, by geforcelegend, GPL-3.0) through Iris 1.8.14. Settings that change from the pack defaults, so that the render is a function of the pose alone and the sun follows the same path as in vanilla:

sunPathRotation=0.0      (the pack tilts the sun 40° by default)
DEPTH_OF_FIELD=false  MOTION_BLUR=false  AVERAGE_EXPOSURE=false
VOLUMETRIC_FOG=false  RAIN_RIPPLES=false  PLANT_WAVE=false  CAULDRON_WAVE=false
CLOUD_SPEED=0.00  PLANE_CLOUD=false  PLANE_CLOUD_SPEED=0.00  WATER_WAVE_SPEED=0.0  CLOUD_TYPE=0

Two captures of the same pose differ by about 43 dB PSNR (temporal accumulation inside the pack): that is the ceiling for any model of this target. The Sundial renders of the second capture round (-r2 shards) are not captured yet. One scene of the first round has no Sundial render.

The renders of Complementary Reimagined used in the accompanying work are not included: its licence does not clearly allow redistributing them.

Splits

Split	Scenes	World
train	16 706	`minesr_atlas`
test	3 381	`minesr_atlas`, in cells of the map not used for training
test	700	`minesr_test2` (swamp), a different world

Biome	Scenes	of them, second round
forest	3 990
plains	3 990	1 190
taiga	3 990	644
desert	3 989	2 589
swamp	3 430
snowy_plains	1 398

World seeds: minesr_atlas 5277617577473417202, minesr_test2 7308878374753708357.

Every position is captured at seven times of day (time_of_day, in game ticks: 0, 1000, 3000, 6000, 9000, 11000, 12000). The split is by area: the map is cut into cells of 512×512 blocks and each plan of poses belongs to one split.

Second round (October 2026). Shards named -r2 hold three new areas: taiga_2, plains_2 and desert_2. The poses of taiga_2_test turned out to lie in the area of the first-round taiga training set (4 blocks from a training pose), so they were moved to train; the second-round test sets are plains_2_test and desert_2_test. Nothing of the first round changed: same scenes, same shards, same split.

Known overlap. Three cells of the first round hold poses of both splits, near the edges of their areas: forest test and swamp train (closest poses 35 blocks apart), plains test and swamp train (32 blocks), and snowy plains train and taiga test (182 blocks). They are kept so that results on this test set stay comparable with published ones.

Animals. Removing entities did not work reliably before mod 0.6.1. In the first round 224 scenes (32 positions) were removed by hand because an animal or a dropped item was in view, and a detector based on depth and the saved world still finds an entity in 3 584 of the 16 364 frames (0.05 % of the pixels). In the second round, taiga_2 and plains_2 were captured with animals (mod 0.6.0); desert_2 with mod 0.6.1, without them.

Files

WebDataset shards, up to 500 scenes and about 1 GB each:

data/<split>/<world>-<biome>[-r2]-<nnn>.tar: the vanilla frame and its annotations;
sundial0/<split>/<same name>.tar: the Sundial Lite render of the scenes of the data shard of the same name.

manifest.jsonl has one line per scene (pose, biome, time, split, shard and sundial0_shard, null when there is no render).

The dataset viewer shows preview/: a 640×360 JPEG thumbnail and the manifest line of every first-round scene. It is only for browsing; the data is in data/ and sundial0/.

Inside a data shard, per scene:

File	Content
`<scene_id>.vanilla.png`	the frame, RGB, 1280×720
`<scene_id>.depth.npz`	`depth` float32 (720, 1280): OpenGL window-space depth in [0, 1], 1 = far plane or sky, origin top left
`<scene_id>.json`	pose, biome, weather, `time_of_day`, and `camera`: `position`, `view_matrix` and `projection_matrix` (4×4, column major)
`<scene_id>.blocks.npz`	`ids` uint16 (720, 1280): index into `names`, the block under each pixel (`nrmc:sky` for the sky)
`<scene_id>.shadow.npz`	sun shadow of each pixel, traced through the blocks of the saved world (see below)

Inside a sundial0 shard: <scene_id>.sundial0.png (RGB, 1280×720) and <scene_id>.sundial0.json (the capture metadata of that pass, with the shader options).

Linearise the depth with camera.projection_matrix; with the view matrix this gives the world position of every pixel.

`shadow.npz`

One ray per pixel from its surface towards the sun, through the block grid of the saved world. This is information the renderer has and the image alone does not: the occluder can be outside the frame.

Array	Content
`bits` uint8	flags of what the ray crossed: 1 solid block, 2 leaves, 4 plant; 8 the first occluder projects inside the frame; 16 it is also visible in the frame
`first` float16	distance in blocks to the first occluder, `inf` if none
`leaves` uint8	number of leaf cells crossed (capped at 255)
`screen_test` uint8	internal flag of the tracer

Loading

Download the shards (hf download stemjara/minecraft-vanilla-frames --repo-type dataset --include "data/*" --include "sundial0/*") and read them with webdataset or tarfile:

import io, json, numpy as np, webdataset as wds

name = "train/minesr_atlas-forest-000.tar"
renders = {s["__key__"]: s["sundial0.png"] for s in wds.WebDataset(f"sundial0/{name}")}
for s in wds.WebDataset(f"data/{name}"):
    meta = json.loads(s["json"])
    depth = np.load(io.BytesIO(s["depth.npz"]))["depth"]
    shadow = np.load(io.BytesIO(s["shadow.npz"]))
    vanilla_png, sundial_png = s["vanilla.png"], renders.get(s["__key__"])
    break

Licence

The images are screenshots of Minecraft, © Mojang AB, shared under Mojang's usage guidelines. This dataset is not an official Minecraft product and is not approved by or associated with Mojang or Microsoft.
The sundial0 renders were made with Sundial Lite v1.2.0, a shader pack under GPL-3.0-or-later. The licence covers the pack's code, not the images it renders; the pack is credited here, and the renders are shared under the same terms as the rest of the images.
Everything derived by the authors (metadata, block maps, shadow traces, manifest) is released under CC BY 4.0.