Performance Benchmarks¶

Compression performance comparisons on standard datasets.

Training Details¶

Unless specified otherwise, models were trained with:

Parameter	Value
Dataset	Vimeo90K
Patch size	256x256
Training steps	4-5M
Batch size	16-32
Initial learning rate	1e-4
LR schedule	ReduceLROnPlateau (patience=20)

Training typically takes 1-2 weeks depending on the model and GPU.

\[\mathcal{L} = \lambda \cdot 255^{2} \cdot \mathcal{D} + \mathcal{R}\]

\[\mathcal{L} = \lambda \cdot (1 - \mathcal{D}) + \mathcal{R}\]

Where \(\mathcal{D}\) is distortion and \(\mathcal{R}\) is estimated bit-rate.

Quality	1	2	3	4	5	6	7	8
MSE	0.0018	0.0035	0.0067	0.0130	0.0250	0.0483	0.0932	0.1800
MS-SSIM	2.40	4.58	8.73	16.64	31.73	60.50	115.37	220.00

Note

MS-SSIM models were fine-tuned from MSE pre-trained networks with learning rate 1e-5.

Bit-rate	Entropy Bottleneck Channels	Recommended
<0.5 bpp	192	Low bit-rates
>0.5 bpp	320	High bit-rates

See tinify.zoo.image.cfgs for detailed configurations.

PSNR on Kodak

Tinify models compared against traditional codecs (JPEG, BPG, VVC/VTM) on the Kodak dataset.

python -m tinify.utils.eval_model pretrained /path/to/kodak/ \
    -a mbt2018-mean -q 1 2 3 4 5 6 7 8

# BPG codec
python -m tinify.utils.bench bpg /path/to/images/

# VTM (VVC reference)
python -m tinify.utils.bench vtm /path/to/images/

python -m tinify.utils.plot results.json --show

For more comparisons and evaluations, see the Tinify paper.