Model Zoo¶

Pre-trained models for end-to-end image and video compression.

Overview¶

Tinify provides pre-trained models optimized with mean square error (MSE) on RGB channels. Models fine-tuned with other metrics are planned for future releases.

Usage¶

Load a pre-trained model:

from tinify.zoo import mbt2018_mean

# Load model with quality level 3
model = mbt2018_mean(quality=3, pretrained=True)
model.eval()
model.update()  # Required for entropy coding

Note

Pre-trained model weights are automatically downloaded to a cache directory. See PyTorch documentation for details.

Input Requirements¶

Requirement	Value
Input shape	(N, 3, H, W)
Minimum H, W	64 pixels
Value range	[0, 1]
Normalization	None (do not normalize)

Warning

Input dimensions may need to be padded to powers of 2, depending on the model architecture.

Train vs Eval Mode¶

Models behave differently in training and evaluation modes (e.g., quantization operations):

model.train()  # Training mode
model.eval()   # Evaluation mode

Available Models¶

Image Compression¶

Model	Quality Levels	Paper	Pre-trained
`bmshj2018_factorized`	1-8	Ballé 2018	Yes
`bmshj2018_hyperprior`	1-8	Ballé 2018	Yes
`mbt2018_mean`	1-8	Minnen 2018	Yes
`mbt2018`	1-8	Minnen 2018	Yes
`cheng2020_anchor`	1-6	Cheng 2020	Yes
`cheng2020_attn`	1-6	Cheng 2020	No

Video Compression¶

Model	Paper	Pre-trained
`ssf2020`	Agustsson 2020	Yes

Model Descriptions¶

bmshj2018_factorized¶

The simplest model with a factorized prior. No hyperprior network.

from tinify.zoo import bmshj2018_factorized
model = bmshj2018_factorized(quality=3, pretrained=True)

bmshj2018_hyperprior¶

Adds a scale hyperprior for better entropy modeling.

from tinify.zoo import bmshj2018_hyperprior
model = bmshj2018_hyperprior(quality=3, pretrained=True)

mbt2018_mean¶

Extends hyperprior with mean prediction for improved compression.

from tinify.zoo import mbt2018_mean
model = mbt2018_mean(quality=3, pretrained=True)

mbt2018¶

Full autoregressive model with joint priors. Best quality but slower.

from tinify.zoo import mbt2018
model = mbt2018(quality=3, pretrained=True)

cheng2020_anchor¶

Attention-based architecture with Gaussian mixture likelihoods.

from tinify.zoo import cheng2020_anchor
model = cheng2020_anchor(quality=3, pretrained=True)

ssf2020 (Video)¶

Scale-space flow model for video compression.

from tinify.zoo import ssf2020
model = ssf2020(quality=3, pretrained=True)

Listing All Models¶

from tinify.zoo import models
print(list(models.keys()))

Cross-Platform Limitations

All models use floating point operations, which are not reproducible across different devices. Encoding/decoding across different platforms is not guaranteed to work. See Integer Networks for Data Compression for solutions.