Cool-chic 5.0: Overfitted Neural Image Codec

• Cool-chic 5.0 is an overfitted image codec that customizes a tiny decoder and hierarchical latent grids per image for targeted rate–distortion optimization.
• It integrates advanced features like inter-feature entropy modeling, hyperlatent grids, and a linear stabilizer to enhance convergence and reduce decoding complexity.
• The codec achieves competitive BD-rate reductions (up to –11.09% vs VVC) with dramatically lower decoder complexity compared to traditional neural autoencoders.

Cool-chic 5.0 is an overfitted image codec in the Cool-chic family: instead of applying a fixed, dataset-trained encoder–decoder pair to arbitrary images, it compresses each image by optimizing a tiny decoder network and a hierarchical latent representation specifically for that image, then transmits both the optimized latents and the decoder parameters. In the literature, the designation has been used in two closely related senses: the perceptually tuned CLIC 2024 candidate, which preserved the original overfitted lightweight decoder while replacing pure MSE with an MSE+MS-SSIM objective and VMAF-based bitstream selection, and the later 2026 codec version that adds inter-feature entropy modeling, hyperlatent grids, linear stabilizer branches, and a faster optimization pipeline (Ladune et al., 2024, Ladune et al., 4 May 2026).

1. Nomenclature and historical position

The Cool-chic lineage originates in "COOL-CHIC: Coordinate-based Low Complexity Hierarchical Image Codec" (Ladune et al., 2022), which introduced a coordinate-based low complexity hierarchical image codec with 629 parameters and 680 multiplications per decoded pixel. That first formulation already established the central design pattern: hierarchical discrete latent grids, a tiny synthesis network, an autoregressive probability model, and per-image optimization rather than dataset-level generalization.

Subsequent work sharpened two distinct meanings of the label "Cool-chic 5.0". In the 2024 CLIC report, Cool-Chic 5.0 denoted a perceptually tuned, challenge-oriented version of the original image codec. It preserved the overfitted lightweight neural decoder jointly optimized with a latent representation per image, but changed the distortion term to combine MSE and MS-SSIM and used VMAF-driven bitstream selection to satisfy the challenge operating points at 0.075, 0.150, and 0.300 bpp (Ladune et al., 2024). In 2026, "Cool-chic 5.0" became the formal name of a new codec generation whose defining changes are inter-feature entropy modeling, hyperlatent grids, a linear stabilizer layer, a refined differentiable quantization proxy, and SOAP for network optimization (Ladune et al., 4 May 2026).

This nomenclature matters because the two usages are not contradictory. The 2024 version is a perceptual operating-point adaptation of the earlier Cool-chic framework, whereas the 2026 version is an explicit architectural and optimization update. A common misconception is therefore to treat "Cool-chic 5.0" as a single immutable specification. The literature instead uses the term for a family state: first as a challenge-tuned perceptual codec, later as a broader architectural revision grounded in faster encoding and stronger entropy modeling (Ladune et al., 2024, Ladune et al., 4 May 2026).

2. Overfitted coding paradigm and signal representation

Cool-chic 5.0 belongs to the class of overfitted codecs. Its operating principle is to optimize, for each image, both the decoder parameters and the latent representation under a rate–distortion objective, rather than to run a generic analysis transform learned offline. In the 2026 formulation, the encoder solves

$$f_{\text{up}}^{*}, f_{\text{syn}}^{*}, f_{\text{ARM}}^{*}, f_{\chi}^{*}, \hat{\mathcal{Y}}^{*} = \arg\min \mathrm{D}(x,\hat{x}) + \lambda \mathrm{R}(\hat{\mathcal{Y}}),$$

with distortion measured by MSE and rate estimated from the learned entropy model (Ladune et al., 4 May 2026).

The main image representation is a set of $L$ discrete latent grids

$$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$

where

$$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$

New in the 2026 version is a second set of hyperlatent grids

$$\hat{\mathcal{Y}}_h = \{ \hat{Y}_h^{(b)}, \ldots, \hat{Y}_h^{(L_h-1)}\},$$

with

$$\hat{Y}_h^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$

These hyperlatents are used only by the entropy model and are discarded once entropy decoding finishes (Ladune et al., 4 May 2026).

The decoder is organized in two stages. First, latent and hyperlatent grids are entropy-decoded with an autoregressive probability model augmented by inter-feature context. Second, the decoded latent pyramid is mapped to the reconstructed image. The image synthesis path begins with a neural upsampling transform

$Z = f_{\text{up}}(\hat{\mathcal{Y}}),$

where $Z$ has shape $(L,H,W)$, followed by an overview transform

$\hat{x} = f_{\text{syn}}(Z),$

with $L$0 RGB output channels (Ladune et al., 4 May 2026).

A distinctive refinement in 5.0 is the linear residual stabilizer. Both the autoregressive entropy model and the synthesis network are split into a nonlinear trunk plus a parallel linear branch,

$L$1

which acts as a residual linear path and improves convergence during per-image optimization (Ladune et al., 4 May 2026).

3. Entropy model, inter-feature context, and optimization procedure

The latent distribution is modeled autoregressively as

$L$2

with each scalar latent assigned a Laplace-distributed likelihood integrated over the quantization bin,

$L$3

The context vector is

$L$4

where $L$5 is the spatial causal context within the current grid and $L$6 is the inter-feature context computed from other already decoded grids (Ladune et al., 4 May 2026).

The main new entropy-model component is the Inter Feature Context Extractor (IFCE). When decoding latent level $L$7, all coarser latent levels and all hyperlatents are already available; they are upscaled by nearest-neighbour to the resolution of level $L$8, stacked into a vector $L$9, and mapped by a small network $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$0 to the inter-feature context

$$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$1

Only the top few resolutions typically use IFCE, namely $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$2, $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$3, and $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$4. Nearest-neighbour upsampling is chosen specifically to avoid multiplications and to ease integer implementations (Ladune et al., 4 May 2026).

The differentiable quantization proxy is also substantially revised in 5.0. During training, hard scalar quantization is replaced by

$$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$5

where $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$6 is a softround operator,

$$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$7

and $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$8 is i.i.d. Gaussian noise (Ladune et al., 4 May 2026). The schedule is explicit: $$\hat{\mathcal{Y}} = \{ \hat{Y}^{(0)}, \ldots, \hat{Y}^{(L-1)}\},$$9 decreases from 0.35 to 0.08, and $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$0 decreases from 0.22 to 0.15 during the main stage.

Optimization is hybridized across parameter classes. SOAP is used for the decoder networks, while Adam optimizes the much larger latent grids. The per-image optimization itself has three stages: a warm-up stage with 400 iterations per candidate over 5 candidates and then 2 candidates, a main stage of 96,700 iterations, and a hardround stage of 500 iterations. The total is therefore 100,000 iterations, with a cosine learning-rate schedule from $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$1 to $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$2, and the rate–distortion trade-offs use

$$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$3

(Ladune et al., 4 May 2026).

4. Decoder complexity, coding gains, and ablation evidence

Cool-chic 5.0 is explicitly designed around a low-complexity decoder. Four operating points are reported for the 2026 codec:

Profile	MAC/pixel	Parameters
LOP	0.5k	0.5k
MOP	1.0k	1.0k
HOP	2.0k	1.9k
VHOP	3.0k	2.7k

The most emphasized configuration is HOP, with 1,991 MAC/pixel and about 1.9k parameters. On CLIC20 professional validation it reaches –11.09% BD-rate vs VVC, while on Kodak it reaches –2.64% BD-rate vs VVC. In the same CLIC20 comparison, MLIC++ reaches –12.03% BD-rate with 816k MAC/pixel, and the paper characterizes Cool-chic 5.0 as competitive with modern autoencoders such as MLIC++ while featuring 250 times lower decoding complexity (Ladune et al., 4 May 2026).

The codec’s strongest practical claim is not only absolute RD performance but the speed at which that performance is approached. Earlier overfitted codecs such as MORIC and LotteryCodec are reported around 104k–110k iterations, at roughly –5.34% and –5.42% BD-rate vs VVC. Cool-chic 5.0 reaches –6.47% already at 10k iterations, –9.30% at 30k, –10.15% at 50k, and –11.09% at 100k. The paper therefore states that it outperforms all overfitted codecs with 10 times less encoding iterations (Ladune et al., 4 May 2026).

The ablation study clarifies where the gains come from. Relative to the full HOP system, replacing SOAP by Adam for the networks yields +2.8% BD-rate; reverting to the older C3 quantization proxy yields +2.7%; removing IFCE yields +3.8%; removing hyperlatents yields +0.5%; and removing the stabilizer yields +0.4%. These results show that the dominant improvements are split between encoder-side optimization changes and the new inter-feature entropy model, rather than coming from a single isolated architectural trick (Ladune et al., 4 May 2026).

A broader complexity context is provided by adjacent work on reduced-complexity overfitted image coding. That study reports a family of decoders from 300 to 2300 MAC/pixel, including a 300 MAC/pixel decoder that still achieves –5.82% BD-rate vs HEVC, and a practical CPU-only decoder requiring about 100 milliseconds to decode an image on CPU (Blard et al., 2024). This anchors Cool-chic 5.0 within a sustained research program on low-complexity neural decoding rather than as an isolated codec instance.

5. Perceptual tuning, reduced-complexity modes, and video generalizations

The earlier challenge-oriented use of the name refers to a perceptual variant of Cool-chic submitted to the 2024 Challenge on Learned Image Compression. Its distortion is

$$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$4

and bitstream selection is driven by VMAF under the challenge budgets. On the validation set, the reported results are: at 0.075 bpp, Avg VMAF 48.305, Worst VMAF 46.128, PSNR 26.533, MS-SSIM 0.927; at 0.150 bpp, Avg VMAF 65.482, Worst VMAF 63.861, PSNR 28.925, MS-SSIM 0.958; and at 0.300 bpp, Avg VMAF 79.005, Worst VMAF 77.577, PSNR 31.686, MS-SSIM 0.976 (Ladune et al., 2024).

The same research line has also explored a Non-Overfitted (N-O) Cool-chic regime, in which a shared analysis transform replaces per-image gradient descent, bringing encoder complexity closer to autoencoders at the cost of RD performance (Blard et al., 2024). HyperCool then builds directly on that N-O framework by using a hypernetwork to generate content-adaptive decoder parameters in a single forward pass. It reports a 4.9% rate reduction over N-O Cool-chic and, with fine-tuning, reaches HEVC-level compression with 60.4% of the encoding cost of the fully overfitted Cool-chic (Borrell-Tatché et al., 23 Sep 2025).

Cool-chic has also been extended to video. "Cool-chic video: Learned video coding with 800 parameters" (Leguay et al., 2024) adds a non-parametric inter-coding module to the image codec and reports 0.9 kilo multiplications per decoded pixel with 800 parameters overall, while supporting both Random Access and Low-Delay P. A subsequent motion-compensation refinement replaces bilinear pixel-wise warping with an 8-tap sinc-based interpolation filter, block size $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$5, and $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$6 fractional positions, producing a rate decrease of more than 10% and lowering motion-related decoding complexity from 391 MAC per pixel to 214 MAC per pixel (Ladune et al., 29 Jul 2025).

A further image-side extension, LANCE, introduces a forward-signaled spatial hyperprior for local entropy adaptation in overfitted image compression. It reports BD-rate reductions over Cool-Chic 4.0 of 1.40% on Kodak and 1.97% on CLIC 2020 at the high end of a 606–1481 MAC/pixel decoder range, and 2.41% and 2.99% at the low end (Benjak et al., 20 May 2026). This suggests that the Cool-chic 5.0 design should be understood less as a frozen endpoint than as a platform for continued work on entropy modeling, encoder acceleration, and locally adaptive priors.

6. Deployment regime, limitations, and open technical questions

The canonical deployment regime for Cool-chic 5.0 is one in which encoding can be expensive but decoding must be cheap. This follows directly from the overfitted formulation: the codec is optimized per image, frequently for thousands or tens of thousands of iterations, but the resulting decoder remains on the order of $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$7 MAC/pixel rather than the $$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$8–$$\hat{Y}^{(k)} \text{ shape is } \left(\tfrac{H}{2^k}, \tfrac{W}{2^k}\right).$$9 MAC/pixel characteristic of many modern autoencoders (Ladune et al., 4 May 2026, Blard et al., 2024). The 2026 paper gives a practical calibration point: on an RTX 4090, 1,000 iterations ≈ 1 minute for a CLIC20 image, so 5,000 iterations ≈ 5 minutes already yields VVC-beating performance (Ladune et al., 4 May 2026).

Several limitations are explicit. First, per-image optimization remains heavy and is unsuitable for strict real-time applications. Second, network bits are not included in the primary optimization objective; they are handled later through post-hoc network quantization and signaling. Third, network overhead can dominate at small image sizes or very low rates: the 2026 analysis reports network parameter costs of roughly 1.5–2.5 kB per image, which is modest for large CLIC20 files but proportionally large for small Kodak images or very small total file sizes (Ladune et al., 4 May 2026). The reduced-complexity study makes the same point from another angle: lowering decoder complexity to 300 MAC/pixel is possible, but the encoder–decoder trade-off remains central, and the authors explicitly propose “slightly overfitted” schemes and meta-learning-inspired variants as future directions (Blard et al., 2024).

Open technical questions therefore cluster around three fronts. The first is encoding acceleration: HyperCool proposes hypernetwork-generated decoder weights and warm-started fine-tuning, while the reduced-complexity work studies shortened training and non-overfitted surrogates (Borrell-Tatché et al., 23 Sep 2025, Blard et al., 2024). The second is entropy-model adaptivity: Cool-chic 5.0 adds IFCE and hyperlatents, and LANCE extends the idea further with a spatial hyperprior for regional adaptation (Ladune et al., 4 May 2026, Benjak et al., 20 May 2026). The third is video generalization under the same low-complexity constraint, where Cool-chic video and its motion-compensation refinement indicate that conventional codec tools such as block-based finite-precision sub-pixel interpolation can be integrated without abandoning the overfitted decoder philosophy (Leguay et al., 2024, Ladune et al., 29 Jul 2025).

Taken together, these works position Cool-chic 5.0 as a mature expression of overfitted neural compression: a codec family in which hierarchical latent grids, autoregressive entropy models, tiny decoder networks, and image-specific optimization are engineered to challenge both conventional transform codecs and much heavier learned autoencoders.