# Media attribution: diffusion-planner-p150 demo renders

Every render shows the **TT output**: the port running on one Tenstorrent Blackhole p150 (tt-nn; ETH dispatch, 1
command queue, 12x10 grid; the pinned numerics of `tt-model.yaml`), called through the Python API
(`DiffusionPlanner.from_pretrained()`, then `model(inputs=...)`), which returns exactly what `POST /predict` returns.
The `*_tt_vs_cpu.png` views and the denoising view put it next to the fp32 CPU reference of the same Autoware network
(the port's torch reference, itself equal to ONNX Runtime on the shipped ONNX files, with the node's pre- and
post-processing), the oracle of the accuracy gates.

What the bird's-eye views draw: the model's own input tensors in the current ego frame (lanes with their bounds, route
lanes, intersection polygons, stop lines, road borders, the 0.1 s neighbour histories), the 8 s ego plan (blue, a dot
every second), the predicted 8 s paths of the neighbours (thin lines in the class colour) and, for nuScenes, the
logged ego path (grey, hollow dots every second). The camera views draw the plan as a vehicle-width ribbon on the
ground plane of the CAM_FRONT key-frame image, and the logged path as a line. Renderer: the research renderer of the
public-data set (`research/diffusion-planner/public_data/scripts/dp_public_render.py`: same panels, palette, legend and
NC label as the CPU-reference renders of `research/diffusion-planner/media/`), driven by the workspace scripts
`logs/diffusion-planner/docs/scripts/{tt_demo_runs,render_tt_media}.py`; not shipped, because the dataset they read is
not.

| file | source | license |
|---|---|---|
| `dp_kashiwanoha_dense_tt_vs_cpu.png` | the shipped sample `code/tt_diffusion_planner/samples/kashiwanoha_dense.npz`: a scripted scene on the Lanelet2 map AutowareFoundation/map-carla-kashiwanoha@0.2.0 | Apache-2.0 (map: Apache-2.0; scene and render: this repo) |
| `dp_kashiwanoha_dense_denoising_steps_tt.png` | the same sample: the ego row of the 11 solver iterates (`~/debug/denoising_steps`) | Apache-2.0 |
| `dp_straight_road_tt_vs_cpu.png` | the shipped sample `straight_road.npz`, a procedural scene of this repo | Apache-2.0 |
| `dp_tt_vs_cpu_agreement_99_scenes.png` | agreement statistics (ego displacement p150 vs CPU, per scene) of the 99 gated scenes: the 2 samples, 5 research scenes (kashiwanoha / CARLA Town10HD maps and procedural roads) and 92 planning instants converted from nuScenes v1.0-mini | a chart of this port's agreement numbers (no dataset content); nuScenes is credited in the image |
| `dp_nuscenes_scene-0061_kf06_bev_tt_NC.png` | nuScenes v1.0-mini scene-0061 (Singapore One-North, mini_train), key-frame 6 (t = 3.0 s): slip road before a left turn, following a van | **CC BY-NC-SA 4.0** + nuScenes Terms of Use (non-commercial) |
| `dp_nuscenes_scene-0061_kf18_bev_tt_NC.png`, `dp_nuscenes_scene-0061_kf18_cam_front_tt_NC.jpg` | scene-0061, key-frame 18 (t = 9.1 s), in the left turn; CAM_FRONT key-frame image | **CC BY-NC-SA 4.0** + nuScenes Terms of Use |
| `dp_nuscenes_scene-0103_kf12_bev_tt_NC.png`, `dp_nuscenes_scene-0103_kf12_cam_front_tt_NC.jpg` | nuScenes mini_val scene-0103 (Boston Seaport), key-frame 12 (t = 6.0 s) | **CC BY-NC-SA 4.0** + nuScenes Terms of Use |
| `dp_nuscenes_scene-0757_kf11_bev_tt_NC.png`, `dp_nuscenes_scene-0757_kf11_cam_front_tt_NC.jpg` | scene-0757 (Boston Seaport, mini_train), key-frame 11 (t = 5.3 s): approach to a signalised intersection with a bus crossing | **CC BY-NC-SA 4.0** + nuScenes Terms of Use |
| `dp_nuscenes_scene-0916_kf11_bev_tt_NC.png` | nuScenes mini_val scene-0916 (Singapore Queenstown, parking lot), key-frame 11 (t = 5.4 s) | **CC BY-NC-SA 4.0** + nuScenes Terms of Use |
| `dp_nuscenes_scene-0061_bev_tt_NC.gif` | scene-0061, every key-frame from t = 3.0 s to 19.2 s (2 Hz, 33 plans), shown at 2 frames/s | **CC BY-NC-SA 4.0** + nuScenes Terms of Use |

What was changed in the nuScenes renders: the map expansion v1.3, CAN bus and annotations converted to the Autoware
Diffusion Planner input tensors (`research/diffusion-planner/public_data/scripts/nuscenes_dp.py`), bird's-eye views
rendered from those tensors and the p150 outputs; camera images resized from 1600x900 to 960x540 with the planned path
drawn and a caption band added. Every image is at most 960 px wide, nothing is cropped or zoomed in on people or
license plates, and the non-commercial label is drawn into every nuScenes image.

## nuScenes (non-commercial)

> Rendered from the nuScenes dataset (v1.0-mini, CAN bus expansion and map expansion v1.3), (c) Motional AD Inc.,
> CC BY-NC-SA 4.0 and the nuScenes Terms of Use (https://www.nuscenes.org/terms-of-use). Non-commercial use only;
> adaptations under the same license. Motional does not endorse this work. Changes: converted to the Autoware Diffusion
> Planner input tensors, bird's-eye-view renders and resized camera images with the planned path drawn. Cite:
> H. Caesar et al., *nuScenes: A Multimodal Dataset for Autonomous Driving*, CVPR 2020.

The `*_NC` files inherit CC BY-NC-SA 4.0: they are labelled "non-commercial" wherever they appear (each image's own
footer says so). No nuScenes data (images, tensors, outputs) is in this repository; only these renders are.

## Kashiwanoha map (Apache-2.0)

`code/tt_diffusion_planner/samples/kashiwanoha_dense.npz` and its renders are derived from the Lanelet2 map
AutowareFoundation/map-carla-kashiwanoha@0.2.0 (`lanelet2_map.osm`, Apache-2.0 per its dataset card; byte-identical to
the test map of autoware_universe `planning/autoware_diffusion_planner`), with an ego, a route and 88 agents scripted by
this port's research tools. `straight_road.npz` uses no external data.