douyeszn's picture
Add files using upload-large-folder tool
fe34e0a verified
|
Raw
History Blame Contribute Delete
2.3 kB
---
license: mit
task_categories:
- image-to-image
tags:
- rendering
- global-illumination
- path-tracing
- renderformer
---
# SparseRenderFormer Cycles references
Path-traced references in [RenderFormer](https://github.com/microsoft/renderformer)'s
exact HDF5 schema, so they load with RenderFormer's own tooling unchanged.
## Why this exists
RenderFormer was trained on two million scenes and sixteen million Cycles
renders, and neither the scenes nor the sampler that produced them was released.
Without a path-traced reference, a modified architecture can only be evaluated
against RenderFormer's own output, which cannot distinguish "matches the
teacher" from "renders correctly".
## Contents
- `cycles/h5/` — 59 scenes, RenderFormer HDF5 schema
- `cycles/renders/` — Blender Cycles references, 4096 adaptive spp with
denoising, 256x256, EXR
- `cycles/scenes/`, `cycles/meshes/` — scene JSON and OBJ, to re-render or modify
- `large_16384/`, `large_32768/` — 60 larger scenes, exported but
**not yet rendered**
## Protocol
Cycles at 4096 adaptive samples per pixel with denoising, at 256x256. This
matches what RenderFormer and RenderFormer++ both report, verified against their
papers rather than assumed.
## Validation
The released `microsoft/renderformer-v1-base` checkpoint scores **31.23 dB**
over 20 of these scenes, against the 29.8 dB RenderFormer reports at this
resolution. Scoring slightly above is expected: these scenes are simpler than
the asset-derived ones their test set is built from.
## Limitations, stated plainly
- 59 scenes is enough for a pipeline check or an ablation, and **not**
enough to train a renderer. A controlled sweep on this data found PSNR moving
4.53 dB purely with training-set size, so treat results fine-tuned on it as
bounded by the data rather than by the method.
- Rendered scenes cap at 3,841 triangles. The larger splits exist but their
references are not rendered yet.
- Images are EXR and genuinely high dynamic range. Read them with a backend that
preserves that: OpenCV's default EXR path clips to 1.0 and drops alpha, which
is silent and ruins training.
## Attribution
Objects derive from Objaverse; see `ATTRIBUTION.md` for per-object licences.
Renders and derived geometry only, never the original meshes.