---
license: cc0-1.0
task_categories:
- text-to-image
tags:
- hdr
- scene-linear
- logc4
- krea2
- lora-training
- ibl
size_categories:
- n<1K
---

# Krea 2 scene-linear HDR — LogC4 training dataset

320 **scene-linear HDR** training images for the Krea 2 HDR LoRA
(see model: [`LAXMAYDAY/krea2-scene-linear-hdr-lora`](https://huggingface.co/LAXMAYDAY/krea2-scene-linear-hdr-lora)
· tools: [`krea2-hdr-toolkit`](https://github.com/HK416-TYPED/krea2-hdr-toolkit)).

Built by re-projecting **CC0 Poly Haven HDRIs** into normal perspective scene images,
exposure-normalizing (mid-gray → 0.18), and **LogC4-encoding** to `[0,1]` so a wide
scene-linear range is carried through an ordinary 8-bit pipeline.

## What each sample is

- `images/<slug>__v<k>.png` — an 8-bit RGB PNG holding the **LogC4 code**. It looks flat/
  grey (normal for log footage); the HDR lives in the encoding. Recover scene-linear
  radiance with **inverse-LogC4** (code `1.0` → `~470×` mid-gray).
- `images/<slug>__v<k>.txt` — caption (trigger phrase `scene-linear HDR, high dynamic range,
  physically based lighting`).
- `manifest.jsonl` — one row per image: `encoding` (LogC4), `anchor` (0.18), `exposure_scale`,
  crop `yaw/pitch/fov`, and decoded HDR stats (`hdr_peak`, `hdr_stops`, `frac_gt1`).
- `manifest_hdr_rich.jsonl` — **the recommended training subset** (229 crops with source
  peak > 2× and ≥ 3 stops). Train on this.

## Decode to scene-linear (EXR)

```python
import numpy as np, imageio.v2 as iio
_A=(2**18-16)/117.45; _B=(1023-95)/1023; _C=95/1023
import math
_S=(7*math.log(2)*2**(7-14*_C/_B))/(_A*_B); _T=(2**(14*(-_C/_B)+6)-64)/_A
def logc4_to_linear(V):            # V in [0,1] -> scene-linear (>>1 for highlights)
    V=np.asarray(V,np.float64)
    return np.where(V>=_C,(2.0**(14*(V-_C)/_B+6)-64)/_A, V*_S+_T)
lin = logc4_to_linear(iio.imread("images/....png").astype(np.float64)/255.0)  # scene-linear HDR
```
(the toolkit's `decode_to_exr.py` writes this straight to 32-bit EXR.)

## Verified

93% of images decode to real HDR (peak > 1.0). Same PNG read naively as SDR peaks at 1.0;
read correctly via inverse-LogC4 it reaches ~470×. The HDR is in the file, log-encoded.

## Notes & limits

- **8-bit is precision, not range**: 256 log-spaced codes cover 0–470× (~17 codes/stop).
  The frozen VAE is ~8-bit effective anyway, so 8-bit LogC4 is adequate.
- Single-plane LogC4 caps peak radiance ~470×; the sun disk clips (true sun magnitude needs
  a multi-plane scheme). ~9% of crops are genuinely low-DR (overcast) — excluded by the
  hdr-rich manifest.
- Source HDRIs are **Poly Haven (CC0)**; this dataset is released CC0.
- Caption caveat: some "visible sun" captions sit on crops where the sun is out of frame
  (a cropping artifact documented in the project report).