README.md

Inference examples

This directory contains 40 numbered, independent inference examples. Every example uses only its public number; source case names and internal paths are intentionally omitted.

Each numbered directory contains:

• first_frame.png: exact 1536x864 generated RGB first frame used by inference.
• prompts/*.txt: the exact rolling long-inference prompts used for the result.
• condition/*.npz: ordered lossless condition chunks.
• metadata.json: frame count, FPS, prompt windows, and condition chunk ranges.

Prompt files are numbered in playback order. The first rolling window spans 124 frames; each later window reuses 34 overlap frames and contributes 90 retained frames. metadata.json records the conditioning and retained half-open frame ranges for every prompt. A negative conditioning start means that the published clip begins partway through an original rolling window; only the recorded retained range belongs to the numbered example.

Each NPZ chunk has two arrays with the same [T, 192, 336] shape. The exact historical Depth representation is declared by depth_format in each example's metadata.json:

• depth: either little-endian float32 metric first-surface depth in metres, or little-endian uint16 fixed-log Gray16 presentation codes.
• semantic_id: uint8, exact class IDs in [0, 11].

For fixed-log examples, condition_preprocessing records the exact mapping and spatial resampling contracts used by inference. Depth is mapped at 672x384 and then reduced with an integer 2x2 BOX mean; Semantic-ID uses a categorical 2x2 mode with the nearest-center tie break. No interpolation is applied to IDs.

The data runs at 24 FPS. Chunks are consecutive and must be concatenated in filename order. They are storage chunks only and introduce no dropped or duplicated frames.

from pathlib import Path
import numpy as np

example = Path("1")
parts = [np.load(path, allow_pickle=False) for path in sorted((example / "condition").glob("*.npz"))]
depth = np.concatenate([part["depth"] for part in parts], axis=0)
semantic_id = np.concatenate([part["semantic_id"] for part in parts], axis=0)
for part in parts:
    part.close()

semantic_id_palette.json documents the class meanings and provides an optional visualization palette. The stored IDs themselves are not palette colors and have not been passed through a lossy video codec.