v3.4.1

🎉 Cool-chic 3.4.1: Up to a 3x encoding speed-up! 🎉

This release only affects the encoding speed of Cool-chic, mostly by leveraging torch.compile to run faster.

• Speed-up GPU Encoding up to 3 times! Check-out the detailed encoding speed-up measurements
• Change encoder configuration files with better medium and slow encoding presets

v3.4.0

🎉 Cool-chic 3.4.0: 30% less complex! 🎉

This release introduces a new upsampling of the latent variables bringing compression gain. It also allows to rely on smaller Auto-Regressive Modules, lowering the overall decoder complexity significantly.

• New and improved latent upsampling module
  • Leverage symmetric and separable convolution kernels to reduce complexity & parameters count
  • Learn two filters per upsampling step instead of one for all upsampling steps
• 1% to 5% rate reduction for the same image quality
• 30% complexity reduction using a smaller Auto-Regressive Module
  • From 2000 MAC / decoded pixel to 1300 MAC / decoded pixel
  • 10% faster decoding speed

Compression performance

	BD-rate of Cool-chic 3.4 vs. [%]						Avg. decoder complexity
	Cheng	ELIC	Cool-chic 3.3	C3	HEVC (HM 16)	VVC (VTM 19)	MAC / pixel	CPU Time [ms]
kodak	-4.2 %	+7.5 %	-0.9 %	-4.3 %	-17.2 %	+3.4 %	1303	74
clic20-pro-valid	-13.2 %	-0.2 %	-0.3 %	-1.3 %	-25.1 %	-2.3 %	1357	354
jvet	/	/	-0.2 %	/	-18.3 %	+18.6 %	1249	143

Kodak

CLIC20 Pro Valid

JVET Class B

v3.3.0

🚀 Cool-chic 3.3: An even faster decoder 🚀

This release improves the speed of Cool-chic CPU-only C decoder and does not affect the compression performance.

Notable improvements:

• Decode a 720p image in 100 ms, 2x faster than Cool-chic 3.2
• Full integerization of the decoder for replicability
• Reduce decoder memory footprint
• Optimized implementation of 3x3 convolutions & fusion of successive 1x1 convolutions

Dataset	Vs. Cool-chic 3.1	Vs. C3, Kim et al.	Vs. HEVC (HM 16.20)	Vs. VVC (VTM 19.1)	Avg decoder MAC / pixel	Avg decoding time [ms]
kodak	- 1.9 %	- 3.4 %	- 16.4 %	+ 4.5 %	1880	96
clic20-pro-valid	- 4.2 %	- 1.0 %	- 24.8 %	- 1.9 %	1907	364
jvet class B	- 7.2 %	/	- 10.8 %	+ 19.5 %	1803	260

v3.2.1

Cool-chic 3.2: Go fast boiii 🔥 🏎️

This release corresponds to the paper Overfitted image coding at reduced complexity, Blard et al., EUSIPCO 24

Notable improvements:

• Fast CPU-only decoder
  • Decode a 512x768 image in 100 ms
  • C API for binary arithmetic coding
• Encoding time reduced by 35%
• Rate reduction of 5% compared to Cool-chic 3.1

Dataset	Vs. Cool-chic 3.1	Vs. C3, Kim et al.	Vs. HEVC (HM 16.20)	Vs. VVC (VTM 19.1)	Avg decoder MAC / pixel	Avg decoding time [ms]
kodak	- 1.9 %	- 3.5 %	- 16.4 %	+ 4.4 %	1880	168
clic20-pro-valid	- 4.3 %	- 1.1 %	- 24.9 %	- 2.0 %	1907	857
jvet class B	- 7.2 %	/	- 10.9 %	+ 19.4 %	1803	485

v3.1

Cool-chic 3.1: now it does video 🎬

This releases is more or less aligned with the paper Cool-chic video: Learned video coding with 800 parameters, DCC 2024, Leguay et al..

Notable improvements:

• Cool-chic can compress videos composed of at least one intra frame followed by 0 to N inter frames ;
• Inter frame can have one reference (P-frame) or two references (B-frame) ;
• Allow to use the usual coding configurations: All Intra, Low-delay P & Random Access.

Detailed results are available on the Cool-chic page

All Intra coding remains similar to Cool-chic 3.0 and is competitive with VVC, while requiring only 2 000 multiplications per decoded pixel.

Low-delay and Random Access offer performance on par with AVC, while requiring only 500 multiplications per decoded pixel.

v3.0

Cool-chic 3.0: A small quantizer change, a giant leap in performance for Cool-chic 🌕 🚀

Improve image compression performance of Cool-chic 2.0 by retaining some of the encoder-side improvements from C3: High-performance and low-complexity neural compression from a single image or video, Kim et al., CVPR 2024.

Notable improvements:

• 17% to 23% rate reduction compared to Cool-chic 2.0 while maintaining identical decoder-side complexity ;
• Better management of the quantization during training thanks to the softround mechanism and the kumaraswamy noise ;
• Overall refactoring of the code for more convenient training.

Detailed results are available on the Cool-chic page.

Bd-rate against different anchors:

Dataset	Vs. Cool-chic 2	Vs. Cool-chic 1	Vs. C3, Kim et al.	Vs. HEVC (HM 16.20)	Vs. VVC (VTM 19.1)	Min decoder complexity [MAC / pixel]	Max decoder complexity [MAC / pixel]	Avg decoder complexity [MAC / pixel]
kodak	- 19.4 %	- 29.1 %	- 1.6 %	- 14.6 %	+ 6.6 %	299	2291	1841
clic20-pro-valid	- 16.8 %	/	+ 3.3 %	- 21.4 %	+ 2.3 %	545	2295	1897
jvet	- 23.0 %	/	/	- 13.7 %	+ 25.4 %	300	2295	1680

v2.0

Cool-chic 2.0: Towards a better decoder 💪

Improve image compression performance of Cool-chic 1.0 by implementing the changes proposed in Low-complexity Overfitted Neural Image Codec, IEEE MMSP 2023, Leguay et al..

Notable improvements:

• 10% rate reduction compared to Cool-chic 1.0 while maintaining identical decoder-side complexity ;
• Several decoder architecture changes: convolution-based synthesis, learned upsampling for the latent variables ;
• Encoder changes: use modified Straight Through Estimator to model the quantization in the final stage of the encoding ;
• Support YUV 420 files as input ;
• Fixed-point arithmetic for the entropy coding part, ensuring cross-platform encoding and decoding.

Detailed BD-rate performance:

Dataset	vs. Cool-chic 1.0	vs. HEVC (HM 16.20)	Avg. Decoder complexity [MAC / pixel]
kodak	-11.7 %	+ 7.1 %	2291
clic20-pro-valid	/	- 4.9 %	2291
jvet	/	+ 11.9 %	2291

v1.0

Cool-chic 1.0: Where it all began 👼

Initial release of COOL-CHIC: Coordinate-based Low Complexity Hierarchical Image Codec, ICCV 2023, Ladune et al.. Cool-chic is a learned image codec, based on overfitting. As such it learns a tiny decoder (around 1 000 parameters) for each image to compress.

Despite its very low decoding complexity, Cool-chic significantly outperforms JPEG and COIN (an other overfitted codec) while being not too far from HEVC.

BD-rate against different anchors (more results available in the paper):

Dataset	vs. COIN	vs. JPEG	vs. HEVC (HM 16.20)	Avg. decoder complexity [MAC / pixel]
kodak	- 65 %	- 55 %	+ 22.2 %	800