KROMA: HDR Imaging & Ontology Matching

• KROMA is a term used for two distinct systems: one for deep HDR chroma compression and another for ontology matching using LLMs and knowledge retrieval.
• In HDR imaging, KROMA employs a U-Net based GAN with a hue-aware loss to efficiently correct out-of-gamut colors, achieving high PSNR, SSIM, and runtime speed.
• For ontology matching, KROMA integrates retrieval-augmented LLM prompts and bisimilarity refinement, outperforming benchmarks with superior F1 scores.

Searching arXiv for the term “KROMA” and the cited papers to disambiguate the topic and ground the article in the relevant literature. KROMA is an overloaded research term that denotes distinct, unrelated frameworks in recent arXiv literature. In computer vision, KROMA refers to a deep learning-based chroma compression framework for tone-mapped HDR images: a GAN-based method that learns to reproduce the conventional chroma-management step quickly and automatically, taking a tone-mapped SDR image and adjusting its colors so that pixels remain within the target display gamut, here sRGB with a D65 white point, while preserving appearance and reducing hue shifts (Milidonis et al., 2024). In semantic interoperability, KROMA stands for “Ontology Matching with Knowledge Retrieval and LLMs,” a retrieval-augmented ontology matching framework that combines knowledge retrieval, LLM-based yes/no matching, bisimilarity-based concept matching, and ontology refinement (Nguyen et al., 18 Jul 2025). The shared name does not indicate a shared method, domain, or lineage.

1. Terminological scope and disambiguation

The term “KROMA” is used for at least two unrelated systems in the cited literature, and it is not interchangeable with several similarly named terms.

Term	Domain	Meaning in the cited paper
KROMA	HDR imaging	Deep chroma compression framework for tone-mapped HDR images (Milidonis et al., 2024)
KROMA	Ontology matching	“Ontology Matching with Knowledge Retrieval and LLMs” (Nguyen et al., 18 Jul 2025)
CHROMA	Image forensics	Detector of AI-generated images through inter-channel color-space correlations (Sotelo et al., 7 Jun 2026)
Chroma+	Stellar modeling	Suite extended for in situ exoplanet transit lightcurve modelling (Short, 2020)
kromatic symmetric function	Algebraic combinatorics	Set-valued $K$-theoretic generalization studied via LC-Hopf algebras (Marberg, 2023)

In (Milidonis et al., 2024), KROMA is not a separate color space or display standard; it is the paper’s proposed deep model for replacing the expensive conventional gamut-management pipeline used after tone mapping. In (Nguyen et al., 18 Jul 2025), KROMA is not related to HDR rendering or color management; it is an ontology matching framework. The neighboring terms “CHROMA,” “Chroma+,” and “kromatic” are likewise distinct: CHROMA addresses AI-image detection (Sotelo et al., 7 Jun 2026), Chroma+ concerns stellar atmosphere and transit modeling (Short, 2020), and the “kromatic symmetric function” belongs to algebraic combinatorics rather than imaging or ontology alignment (Marberg, 2023).

A common misconception is that the shared orthography implies a single technical concept. The cited papers do not support that reading. The name instead labels separate systems in different subfields.

2. KROMA in HDR imaging: problem setting and conceptual role

In HDR-to-SDR rendering, the KROMA of (Milidonis et al., 2024) addresses a specific failure mode of tone mapping. The paper emphasizes that tone mapping primarily compresses luminance, not chroma. As a result, after tone mapping, many pixels can still fall outside the display gamut. If such pixels are shown on an SDR display, they are clipped or forced back into gamut, which can cause visible chromatic distortions, over-saturated regions, desaturation, and hue shifts (Milidonis et al., 2024).

Previous studies suggested that a gamut management step ensures that all pixels remain within the target gamut, but such approaches are computationally expensive and cannot be deployed on devices with limited computational resources. The paper positions KROMA as a fully data-driven and feed-forward alternative to conventional gamut mapping or chroma compression. Traditional methods compute corrections by analytically modeling the gamut boundary and processing lightness/chroma directly, often requiring both the tone-mapped image and the original HDR image or multiple manual stages. By contrast, KROMA learns this correction end-to-end from examples, needs only the tone-mapped image at inference time, and is compatible with multiple TMOs (Milidonis et al., 2024).

The reference target gamut is explicitly sRGB with a D65 white point. The problem formulation is a paired image-to-image task: the network learns a mapping from a tone-mapped image to its corresponding chroma-compressed result. The paper’s own framing stresses that it is the first deep learning model dedicated specifically to chroma compression, whereas prior deep learning work in HDR mostly focused on tone mapping, inverse tone mapping, or enhancement rather than gamut-constrained chroma correction (Milidonis et al., 2024).

This suggests that the paper treats chroma compression not as a secondary post-processing heuristic but as a distinct learnable operator within the HDR display pipeline.

3. Network architecture, training protocol, and hue-aware objective

The final KROMA model in (Milidonis et al., 2024) is a paired image-to-image GAN. After comparing a GAN architecture with other image translation and enhancement networks, the paper adopts a U-Net generator, a pixel-based discriminator rather than a patch discriminator, a least-squares GAN (LSGAN) objective for training stability, instance normalization, and orthogonal weight initialization. The paper notes that the pixel-based discriminator helps suppress tiling artifacts that can appear with patch-based discriminators, while preserving detail. The generator and discriminator use 128 and 64 filters, respectively, in the specified terminal/initial convolutional layers. Training uses Adam with momentum $0.5$, learning rate $10^{-4}$, linear decay, and 50 decay iterations. Importantly, the model is trained without pre-processing so the images remain at native resolution (Milidonis et al., 2024).

A central contribution is the hue-aware loss. The paper states that conventional $L1$ or $L2$ losses alone do not sufficiently penalize chromatic inaccuracies in perceptually important hue regions. The hue loss is defined as

$$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$

where $I_H$ is the hue channel of the reference image and $\tilde{I}_H$ is the hue channel of the generated image. The full objective is

$$G^*=\arg\min_{G}\max_{D}\mathcal{L}_{cGAN}(G,D)+\lambda_1\mathcal{L}_{L1}(G)+\lambda_H\mathcal{L}_{H}(G)$$

with optimal weights found empirically as $\lambda_1 = 100$ and $0.5$0 (Milidonis et al., 2024).

The stated purpose of this loss combination is to balance global realism from the adversarial term, local fidelity from $0.5$1, and hue fidelity from the hue loss. The authors explicitly state that the hue term improves chromatic accuracy, especially in high-hue or saturated regions. This suggests that KROMA’s design is motivated less by generic image enhancement and more by the specific geometry of hue preservation under gamut constraints.

4. Dataset, evaluation protocol, and reported performance

The training corpus for the HDR-imaging KROMA is substantial. The authors assembled 1467 HDR images at $0.5$2 resolution from their own captures and several HDR repositories. These HDR images were tone-mapped using a wide variety of TMOs, then chroma-compressed using the conventional reference pipeline. This produced 42,543 paired examples, split 8:1:1 into training, validation, and test sets $0.5$3. By using many TMOs rather than only one, the dataset was designed to help the network generalize across different tone-mapping styles. The paper also reports leave-one-out TMO experiments to test whether the model remains effective when confronted with tone-mapped images from unseen operators (Milidonis et al., 2024).

Evaluation uses PSNR, SSIM, $0.5$4 as mean absolute error of the hue channel in CIELCh, $0.5$5 as color difference in CIELab, and CVVDP reported in JOD units. Quantitative results show the proposed method outperforming compared baselines including Pix2pix, CycleGAN, Pix2pixHD, and HDRNet. The reported scores for the proposed model are PSNR $0.5$6 dB, SSIM $0.5$7, $0.5$8, $0.5$9, and CVVDP $10^{-4}$0 JOD. HDRNet, described as the strongest baseline in the summary, scores $10^{-4}$1 dB / $10^{-4}$2 / $10^{-4}$3 / $10^{-4}$4 / $10^{-4}$5 (Milidonis et al., 2024).

Across individual TMOs, performance remains strong, with PSNR ranging from $10^{-4}$6 to $10^{-4}$7 dB, SSIM from $10^{-4}$8 to $10^{-4}$9, and CVVDP near the maximum possible value. The leave-one-out results show that the model generalizes well to TMOs it was not trained on (Milidonis et al., 2024).

The subjective study uses a 2AFC design with 21 participants who had normal or corrected vision and no color blindness. Ten test images spanning indoor/outdoor and daytime/nighttime scenes were used, with 100 total comparisons. Viewing was performed on an Eizo ColorEdge CG319X 4K HDR calibrated monitor in a dark room, set to SDR mode, sRGB, D65, and $L1$0 peak luminance. Participants chose the image they found more realistic/natural and of higher overall visual quality. The paper concludes that the proposed network is equivalent to Pouli’s and Sikudov’s analytic methods in visual quality, while being selected significantly more often than Artusi’s and Mantiuk’s methods (Milidonis et al., 2024).

5. Runtime characteristics and deployment implications of the HDR-imaging KROMA

Runtime is a major part of the KROMA claim in (Milidonis et al., 2024). The abstract states that the method is up to 3 orders of magnitude faster than the conventional chroma-compression pipeline, and the detailed timing data substantiate a large speed advantage. At $L1$1 resolution, the model takes about 15 ms per image, versus 627 ms for the conventional CUDA version and 3.8 s for the CPU version, roughly $L1$2 and $L1$3 faster, respectively. It maintains real-time performance up to 1 megapixel at 25 fps, and at 9.4 megapixels it still processes an image in 375 ms, compared with 7.1 s for the CUDA reference and 133 s for CPU (Milidonis et al., 2024).

The paper interprets these timings as evidence that the method is practical for deployment on resource-limited devices such as smartphones and tablets, especially when combined with any existing TMO. The broader implication stated in the summary is that KROMA-like deep chroma compression can make HDR display pipelines much more efficient without sacrificing the color fidelity that conventional tone mapping alone cannot guarantee (Milidonis et al., 2024).

A plausible implication is that the learned replacement of an analytic gamut-management stage may be especially attractive in end-user imaging stacks where latency and power budget are as important as colorimetric fidelity.

6. KROMA in ontology matching: retrieval-augmented LLM alignment and bisimilarity refinement

A second, unrelated KROMA appears in (Nguyen et al., 18 Jul 2025). Here KROMA stands for “Ontology Matching with Knowledge Retrieval and LLMs.” It is a framework for ontology matching, defined as finding semantically equivalent concepts across two ontologies, by combining retrieval-augmented generation to enrich each concept with external and structural context, LLM-based yes/no matching to decide candidate equivalences, and bisimilarity-based refinement to enforce structural consistency and compress the final aligned ontology into a minimal concept graph (Nguyen et al., 18 Jul 2025).

The ontology model is a DAG $L1$4. For source and target ontologies,

$L1$5

A matching induces an ontological equivalence relation $L1$6, which partitions $L1$7 into equivalence classes, and the induced quotient structure is the concept graph $L1$8. The paper proves two properties summarized as follows: the ontological bisimilar relation $L1$9 is an equivalence relation, and for a given configuration $L2$0, there exists a unique smallest concept graph induced by $L2$1 (Nguyen et al., 18 Jul 2025).

The end-to-end pipeline includes concept graph initialization, knowledge retrieval, embedding generation and candidate reduction, LLM prompting, bisimilarity-based concept matching, and ontology refinement. Knowledge retrieval gathers structural knowledge by neighborhood sampling over the current concept graph up to two hops, lexical/contextual knowledge by converting the sampled subgraph into SPARQL-like query templates centered on the concept, and definitional/external knowledge by executing those queries against external knowledge sources or knowledge graphs. Embeddings are then formed as

$L2$2

and top-$L2$3 candidate pairs are sent to the LLM rather than exhaustively querying all pairs (Nguyen et al., 18 Jul 2025).

The Boolean matching function is

$L2$4

where $L2$5 is a semantic similarity measure, $L2$6 is the answer from the LLM(s) to the prompt “Are these semantically equivalent?”, and $L2$7 is a configurable weight. Bisimilarity requires a concept similarity threshold together with backward and forward neighbor conditions, so two concepts are equivalent only if they are semantically similar and their local neighborhoods are recursively consistent (Nguyen et al., 18 Jul 2025).

KROMA supports both OfflineRefine and OnlineRefine. The offline algorithm works bottom-up by rank, grouping nodes into buckets $L2$8 and using a collapse operator to split or merge fragments until the partition is consistent with bisimilarity. The offline refinement is claimed to run in $L2$9 for the graph construction/refinement structure, while the online update delay for each batch is

$$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$0

The overall workflow cost is given as

$$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$1

with $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$2 the unit cost of one LLM prompt and $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$3 the size of the external knowledge graph used for retrieval (Nguyen et al., 18 Jul 2025).

7. Empirical results of the ontology-matching KROMA and cross-domain significance of the name

The ontology-matching KROMA is evaluated on five OAEI-style benchmark tracks: Anatomy — Mouse-Human, Bio-LLM — NCIT-DOID, CommonKG — Nell-DBpedia and YAGO-Wikidata, BioDiv — ENVO-SWEET, and MSE — MI-MatOnto. The paper tests DeepSeek-R1-Distill-Qwen-1.5B, GPT-4o-mini, Llama-3.3, Gemma-2B, Llama-3.2-3B, Mistral-7B, and Llama-2-7B, with no fine-tuning, temperature $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$4, fixed random seed, and confidence threshold $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$5 (Nguyen et al., 18 Jul 2025).

For each dataset, 20 matched concept pairs were sampled from ground truth, 20 unmatched source concepts were also sampled, each source concept was paired with 25 candidate target concepts, and this yielded 1000 concept pairs per configuration. Evaluation uses Precision, Recall, and $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$6. The full KROMA configuration KL3.3 achieves the best $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$7 on all six reported tasks: Mouse-Human $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$8, NCIT-DOID $$\mathcal{L}_{H}(G)=\|I_H-\tilde{I}_H\|_2^2 \text{,}$$9, Nell-DBpedia $I_H$0, YAGO-Wikidata $I_H$1, ENVO-SWEET $I_H$2, and MI-MatOnto $I_H$3. The paper states that KROMA outperforms existing LLM-based methods by $I_H$4 on average (Nguyen et al., 18 Jul 2025).

Ablations attribute the gains to both ontology refinement and knowledge retrieval. Removing refinement reduces $I_H$5 from $I_H$6 to $I_H$7 on Mouse-Human, from $I_H$8 to $I_H$9 on NCIT-DOID, from $\tilde{I}_H$0 to $\tilde{I}_H$1 on Nell-DBpedia, from $\tilde{I}_H$2 to $\tilde{I}_H$3 on YAGO-Wikidata, from $\tilde{I}_H$4 to $\tilde{I}_H$5 on ENVO-SWEET, and from $\tilde{I}_H$6 to $\tilde{I}_H$7 on MI-MatOnto. Removing retrieval hurts more substantially, with corresponding drops to $\tilde{I}_H$8, $\tilde{I}_H$9, $$G^*=\arg\min_{G}\max_{D}\mathcal{L}_{cGAN}(G,D)+\lambda_1\mathcal{L}_{L1}(G)+\lambda_H\mathcal{L}_{H}(G)$$0, $$G^*=\arg\min_{G}\max_{D}\mathcal{L}_{cGAN}(G,D)+\lambda_1\mathcal{L}_{L1}(G)+\lambda_H\mathcal{L}_{H}(G)$$1, $$G^*=\arg\min_{G}\max_{D}\mathcal{L}_{cGAN}(G,D)+\lambda_1\mathcal{L}_{L1}(G)+\lambda_H\mathcal{L}_{H}(G)$$2, and $$G^*=\arg\min_{G}\max_{D}\mathcal{L}_{cGAN}(G,D)+\lambda_1\mathcal{L}_{L1}(G)+\lambda_H\mathcal{L}_{H}(G)$$3 (Nguyen et al., 18 Jul 2025).

The paper also studies debating and deep reasoning. Multi-agent debate can help in some settings, but it is expensive and not consistently better; longer debates help more than adding more agents. Forcing long reasoning increased prompt/output size substantially but tended to reduce recall and slightly hurt overall $$G^*=\arg\min_{G}\max_{D}\mathcal{L}_{cGAN}(G,D)+\lambda_1\mathcal{L}_{L1}(G)+\lambda_H\mathcal{L}_{H}(G)$$4. The paper concludes that verbose reasoning may improve transparency but can consume context and over-filter valid matches in binary ontology matching tasks (Nguyen et al., 18 Jul 2025).

Taken together, the two KROMA systems illustrate that the same research label has been used for two very different technical programs: one centered on feed-forward chroma compression in HDR display pipelines, the other on RAG-augmented, bisimilarity-constrained ontology alignment. The literature also makes clear what KROMA is not: it is not CHROMA’s color-correlation forensic detector (Sotelo et al., 7 Jun 2026), not the Chroma+ stellar-modelling suite (Short, 2020), and not the kromatic symmetric function of algebraic combinatorics (Marberg, 2023).