RenderFlow: Single-Step Neural Rendering via Flow Matching
Paper • 2601.06928 • Published
Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
RenderFlow-mini dataset
A 256×256 G-buffer rendering dataset generated with Mitsuba 3.9.1 (cuda_ad_rgb) for the
scaled reproduction of RenderFlow, mirroring the paper's
procedural pipeline: random primitive scenes (spheres / cubes / cylinders on a ground plane,
2–5 objects, random principled materials with Disney-style roughness/metallic/specular), lit by
10 CC0 1k HDRI environment maps from Poly Haven, viewed by orbital cameras.
- 360 sequences × 5 frames (300 train / 60 test, disjoint seeds) — 1,800 frames
- Per frame:
gt(192-SPP path-traced RGB, clipped to [0,1]),albedo(unlit base color, emission pass — exact),aux4-channel buffer = [depth, sh_normal×3] - One
.npzper sequence + per-sequence metadata JSON (object/material params, envmap file, camera) - Full dataset in bucket
mewvan/renderflow-mini(dataset/); this repo mirrors 8 sample sequences (samples/) and preview strips GT|albedo|normals|depth (previews/)
Not a substitute for the paper's 130k-frame UE5 dataset — primitives only, no texture variation,
no emissive objects. Generated by v-file://gen_data.py (see bucket report).
- Downloads last month
- 61