# RecGen — sparse-structure denoiser, finetuned (70k)

An updated **sparse-structure (SS) denoiser** for RecGen. It is a drop-in
replacement for the `stereo_denoiser_ema0.9999_step0055000.pt` at the repository
root; the SLAT denoiser is **unchanged** and should still be taken from the root.

```
stereo_denoiser_ema0.9999_step0070000.pt   this checkpoint
stereo_config.json                         required — architecture + pose representation
stereo_pose_stats.json                     required — pose normalization statistics
```

## Usage

Pair it with the SLAT denoiser at the repository root:

```python
from recgen_utils import load_pipeline
pipeline = load_pipeline(
    checkpoint_slat="slat_denoiser_ema0.9999_step0075000.pt",        # repo root, unchanged
    checkpoint_sparse="sparse-structure-ft-70k/stereo_denoiser_ema0.9999_step0070000.pt",
)
```

## What changed

Finetuned from the released SS checkpoint on a preprocessed multi-view dataset
with SAM2-mixed masks and per-sample color jitter.

| | |
|---|---|
| initialized from | released `stereo_denoiser_ema0.9999_step0055000.pt` |
| optimizer | AdamW, lr 1e-5, weight decay 0 |
| augmentation | color jitter (brightness/contrast/saturation 0.4, hue 0.1) |
| masks | mixed source, 50% SAM2 |
| EMA | 0.9999 |
| architecture | unchanged — 1024 channels, 2 pose tokens, `6d_translation_scale`, 2 views @ 518px |

## Evaluation

Full HomebrewedDB evaluation, 539 samples, **single-view conditioning**, with the
SLAT denoiser held fixed at the released checkpoint so that differences are
attributable to the sparse-structure stage alone.

| metric | released SS @55k | this checkpoint | |
|---|---|---|---|
| ADD-S recall @2% of diameter | 0.048 | **0.093** | ~1.9× |
| Chamfer distance, median (cm) | 0.410 | **0.403** | |
| Chamfer distance, 90th pct (cm) | 0.800 | **0.740** | −7.5% |
| Chamfer distance, pre-ICP median (cm) | 0.669 | **0.627** | −6.3% |

ADD-S recall at the conventional 10%-of-diameter threshold is saturated for both
checkpoints (0.950 vs 0.961) and does not separate them; the tighter 2% threshold
is where the difference appears. The improvement is concentrated in pose accuracy
and in the hard tail rather than in median reconstruction accuracy.

## License

CC-BY-NC-4.0, as the rest of this repository.