README
RightWayUp training renders
The 3,586 Blender renders that ORTUS AI made from Poly Haven assets to train RightWayUp, plus the 791 renders used in its held-out clean scene test. Every image has a known roll, taken from the Blender camera.
Model · Code and technical report · Video and write-up
Contents
| Path | Files | What |
|---|---|---|
renders/ | 3,466 PNG + 3,466 JSON | Upright renders of 15 scene types (carpark, corridor, healthcare, hospitality, industrial yard, loading bay, lobby, office, outdoor, residential, retail, school, stairwell, station platform, warehouse) |
experiments/ | 120 JPG + 120 JSON | Renders with a known non-zero roll from a direct camera-perspective set |
evaluation/clean-v1-frozen/ | 791 PNG | The render images of RightWayUp's held-out clean scene test |
index.jsonl | 3,586 rows | One row per training render: file, sha256, base_roll_cw (clockwise roll of the image content, degrees), training loss weight, supervision, Poly Haven assets |
SHA256SUMS | Hash of every file |
Each JSON sidecar records how the image was made: Blender version and render engine, camera matrix and attitude, the projected gravity direction, and the Poly Haven asset files with their hashes.
Every render row in the model's per-image training manifests (manifests/training/gf_poly_haven.jsonl.gz and
gf_poly_direct.jsonl.gz in ortusai/rightwayup)
gives the file's path in this dataset (locator.render_file) and its SHA-256 (locator.render_sha256). Training used
resized lossless copies of these files; the manifests record each copy's hash (stored_sha256). The evaluation renders are listed the same
way in manifests/evaluation/clean-v1-frozen-parents.jsonl in the
GitHub repository, which also has the view
definitions (angles, crops, degradations) to rebuild the test.
from huggingface_hub import snapshot_download
path = snapshot_download("ortusai/rightwayup-renders", repo_type="dataset")
Licence
The renders are © 2026 ORTUS AI SOFTWARE LIMITED and released under
CC BY 4.0 (LICENSE). The licence covers our renders; the Poly Haven
assets in them are CC0 and stay CC0 (credit to Poly Haven is appreciated, not required). Please credit
"RightWayUp training renders by ORTUS AI" with a link to this page.
Similar repos
README
RightWayUp training renders
The 3,586 Blender renders that ORTUS AI made from Poly Haven assets to train RightWayUp, plus the 791 renders used in its held-out clean scene test. Every image has a known roll, taken from the Blender camera.
Model · Code and technical report · Video and write-up
Contents
| Path | Files | What |
|---|---|---|
renders/ | 3,466 PNG + 3,466 JSON | Upright renders of 15 scene types (carpark, corridor, healthcare, hospitality, industrial yard, loading bay, lobby, office, outdoor, residential, retail, school, stairwell, station platform, warehouse) |
experiments/ | 120 JPG + 120 JSON | Renders with a known non-zero roll from a direct camera-perspective set |
evaluation/clean-v1-frozen/ | 791 PNG | The render images of RightWayUp's held-out clean scene test |
index.jsonl | 3,586 rows | One row per training render: file, sha256, base_roll_cw (clockwise roll of the image content, degrees), training loss weight, supervision, Poly Haven assets |
SHA256SUMS | Hash of every file |
Each JSON sidecar records how the image was made: Blender version and render engine, camera matrix and attitude, the projected gravity direction, and the Poly Haven asset files with their hashes.
Every render row in the model's per-image training manifests (manifests/training/gf_poly_haven.jsonl.gz and
gf_poly_direct.jsonl.gz in ortusai/rightwayup)
gives the file's path in this dataset (locator.render_file) and its SHA-256 (locator.render_sha256). Training used
resized lossless copies of these files; the manifests record each copy's hash (stored_sha256). The evaluation renders are listed the same
way in manifests/evaluation/clean-v1-frozen-parents.jsonl in the
GitHub repository, which also has the view
definitions (angles, crops, degradations) to rebuild the test.
from huggingface_hub import snapshot_download
path = snapshot_download("ortusai/rightwayup-renders", repo_type="dataset")
Licence
The renders are © 2026 ORTUS AI SOFTWARE LIMITED and released under
CC BY 4.0 (LICENSE). The licence covers our renders; the Poly Haven
assets in them are CC0 and stay CC0 (credit to Poly Haven is appreciated, not required). Please credit
"RightWayUp training renders by ORTUS AI" with a link to this page.