Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.28964

Can Quantum Layers offer Performance and Security Advantages in Vision-Language Models?

Md Khalid Syfullah, Alvi Ataur Khalil

quant-phcs.AIcs.CVcs.LGSubmitted Oct 20, 2026• Announced Oct 22, 2026• 0 citations
arXivHTMLPDF

Abstract

As quantum computing moves from the noisy intermediate-scale quantum (NISQ) era toward early fault-tolerant quantum computing (EFTQC), quantum-enhanced AI is shifting from a conceptual idea to experimental research. Quantum components are expected to extend the capabilities of existing AI systems, yet it remains unclear how such components will help within large multimodal models. In this study, we try to answer that question for vision-language models (VLMs) using classically simulated quantum layers. With LLaVA-OneVision as the backbone, we evaluate 4- and 8-qubit basic-ring and strongly entangling circuits at four insertion and four replacement locations. Insertion adds quantum layers to the classical model, while replacement substitutes them for classical components. Using the same setup, we compare performance and computational cost between classical and quantum-classical hybrid models. On the Visual Spatial Reasoning dataset, inserting a quantum layer after the visual projector reaches 80.7% development accuracy, 2.28 percentage-points above classical full fine-tuning. Replacing the final decoder feed-forward sublayer retains 79.22% accuracy with 13.07 million fewer trainable parameters. Quantum insertion can thus improve accuracy, and selective replacement can reduce parameters. These benefits depend strongly on where the quantum layer is inserted or which classical component it replaces: earlier replacements lower accuracy, and fewer parameters do not reliably shorten training or reduce computation. To assess potential security benefits, we evaluated the robustness of quantum-classical hybrid models under six attack settings. Overall, our study maps where quantum layers help and hurt within VLMs, offering a starting point for hybrid VLM design; whether these gains generalize remains open. The code is available at https://github.com/TITANS-Lab-Artifacts/QuantumVLM.

AI Summary

The paper asks whether small quantum layers can improve vision-language model accuracy, reduce computational cost, or increase adversarial robustness. It evaluates noiseless, classically simulated 4- and 8-qubit variational circuits in LLaVA-OneVision-Qwen2-0.5B, comparing two circuit families at four insertion and four component-replacement locations. Experiments use the Visual Spatial Reasoning dataset, with 7,680 training and 1,097 development examples, alongside classical fine-tuning and parameter-efficient baselines. Under full-backbone training, inserting an 8-qubit strongly entangling circuit after the visual projector achieves 80.7% development accuracy, 2.28 percentage points above classical full fine-tuning; classical LoRA reaches 79.0%. Replacing the final decoder feed-forward sublayer achieves 79.22% while removing 13.07 million trainable parameters from approximately 893.68 million. Replacing the projector or an early decoder feed-forward sublayer instead yields approximately 51–55% accuracy. Parameter reductions do not consistently reduce measured training time or memory, and unequal early-stopping budgets prevent speedup conclusions. Six image-perturbation attacks are evaluated on 40 development examples at $\ell_\infty$ radius $4/255$. Both iterative white-box PGD attacks reduce robust accuracy to zero for every model; no consistent robustness benefit is observed. The stated limitations include one backbone, dataset, and training seed; development-set reuse for checkpoint selection and reporting; selection among many configurations; absent parameter-matched classical controls; and unequal training budgets. Robustness interpretation is further limited by the small subset, HSJ initialization failures, and unresolved gradient diagnostics. No quantum-hardware performance or quantum advantage is established.

Little sign of AI-generated textIncrementalSmall numerics onlyMissing comparison

Central claim may not hold: The accuracy gain is selected and reported on the same development split used for checkpoint selection, with only one training seed, so it does not independently establish improved predictive performance. Without matched classical bottleneck or adapter controls, the observed accuracy and parameter-count results do not isolate a benefit attributable specifically to the quantum circuit. Classical and hybrid full fine-tuning use different early-stopping patience values, preventing a matched-budget interpretation of their training-time comparison.

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Related papers

A Scalable Hybrid Quantum-Classical CNN Architecture for High-Performance, Efficient Large-Scale Image Recognition

Krishna Munishamaiah

Table1_Performance.csv This file contains the classification performance of the Classical CNN and the proposed SH-QC-CNN across three benchmark datasets: ImageNet-1K, CIFAR-100, and Medical X-ray. It reports Top-1 accuracy for both models and the percentage improvement achieved by the hybrid architecture. The data supports the claim that quantum-enhanced feature processing consistently improves recognition accuracy across diverse domains. Fig7_Parameter_Counts.csv This dataset reports the number of trainable parameters (in millions) for Classical CNNs and SH-QC-CNN on three datasets. It demonstrates the substantial reduction in model size achieved by replacing dense classical layers with Parameterized Quantum Circuits (PQCs), highlighting the parameter efficiency of the proposed hybrid architecture. Fig8_Training_Time.csv This file provides the total training time (in hours) for Classical CNNs and SH-QC-CNN across ImageNet-1K, CIFAR-100, and Medical X-ray datasets. The data shows that SH-QC-CNN reduces training time by approximately 34% on average, validating the computational efficiency of quantum-enhanced feature compression. Fig9_PQC_Depth_vs_Accuracy.csv This dataset captures the relationship between PQC depth (number of quantum layers) and classification accuracy. It illustrates that performance improves with depth up to three layers, after which accuracy degrades due to NISQ-era limitations such as noise and barren plateaus. The data justifies the selection of a three-layer PQC as the optimal configuration. Fig11_Noise_Robustness.csv This file contains clean and noisy accuracy values for both Classical CNN and SH-QC-CNN under Gaussian noise perturbation. It quantifies the accuracy drop and demonstrates that SH-QC-CNN is significantly more robust to input noise, confirming the stabilizing effect of quantum feature embeddings. Fig12_Qubit_Scaling.csv This dataset shows the relationship between classical feature dimensionality and the number of qubits required for quantum encoding. It evidences a logarithmic scaling behavior, indicating that high-dimensional feature spaces can be efficiently represented using a small number of qubits, making the architecture feasible for NISQ-era hardware.

•Jan 23, 2026•0 citations

Hybrid Quantum-Classical Neural Networks: Bridging Deep Learning and Quantum Computing for Image Classification

Akash Jha, Rahul Bhandari, Raman Chadha

Given the recent prominence of Noisy Intermediate Scale Quantum (NISQ) devices, hybrid quantum-classical neural networks have recently drawn a lot of excitement as a potential solution to improving learning efficiency. Our current paper describes a Hybrid Quantum-Classical Convolutional Neural Network (HQCNN) which is defined as a pretrained backbone of ResNet-18 and a small quantum-enhanced feature-transformation layer of Qiskit. The convolutional high-dimensional features are reduced to the classical eight-dimensional one, which in turn is encoded into a four-qubit parameterized quantum circuit. The resultant quantum expectation values are combined with the compressed features to come up with a hybridized representation of twelve dimensions, used to classify multi-class dataset. Experimental analysis of CIFAR10 data has shown that the suggested HQCNN achieves a test accuracy of 94.74% and improves class-wise precision, recall, and F1-scores than a number of existing hybrid models while using a significantly smaller number of latent features. More investigation into the quantum module indicates that it is predictable about scalability to batch, resistant to perturbations in features, and efficiently performed by evaluating circuit vectors. These results indicate that compact and expressive representations of images that can be helpful in the NISQ era can be acquired using hybrid architectures, thus providing a practical way to forward the benefits of quantum to modern deep-learning pipelines.

International Conference on Emerging Computational Intelligence (ICECI)•Apr 4, 2026•0 citations

Quantum assistant: Specialization of multimodal vision-language models for quantum computing

Samuel Lima Braz, João Paulo Leite

Code assistants for quantum computing operate exclusively on text, disregarding visual representations (circuit diagrams, Bloch spheres, and histograms) that permeate the domain’s documentation. This work is presented as a resource-and-pilot study with three contributions: (1) the first public multimodal dataset for Qiskit-oriented quantum programming, containing 8,366 samples (45% with images) across function completion, code generation, and question answering tasks, released under Apache 2.0; (2) a reproducible synthetic generation pipeline that extracts content from public Qiskit documentation, transcribes images via a multimodal model, generates question-answer pairs, and validates code through automatically generated unit tests; and (3) a feasibility study of parameter-efficient VLM specialization, in which rsLoRA (rank 32, two epochs) on Qwen3-VL-8B-Instruct is retained as the best configuration under a controlled method-selection comparison of five LoRA variants. On the text-only external benchmarks Qiskit HumanEval and Qiskit HumanEval Hard, which measure quantum-code specialization rather than visual understanding, the fine-tuned model reaches 41.72% and 30.46% Pass@1 respectively, above the IBM-specialized Qwen2.5-Coder-14B-Qiskit (40.40% / 25.83%) and granite-3.3-8b-qiskit (17.88% / 9.93%), and + 17.22 / + 15.89 pp over the Qwen3-VL-8B-Instruct base. Visual capability is assessed on the synthetic multimodal test split, where multimodal samples reach 64.29% Pass@1 versus 46.33% on text-only samples ( + 17.96 pp); a matched training ablation that withholds the multimodal portion of the training set localizes + 6.25 pp of that gap to the multimodal training signal. The dataset, models, and code are available under the Apache 2.0 license.

Expert Systems with Applications•Aug 13, 2026•1 citation

QuantumGPTMini: A Hybrid Quantum-Classical Transformer for Enhanced NLP

Ramazan Amire, Amire Ramazan

Transformer architectures have driven advances in natural language processing, powering systems from question answering to creative text generation. Yet as model sizes have grown from hundreds of millions to hundreds of billions of parameters, the computational cost of these systems has become prohibitive, placing advanced language technology out of reach for many researchers. In this work, we introduce QuantumGPTMini, a hybrid quantum-classical transformer with just over one million classical parameters that leverages lightweight quantum circuits to perform its core attention and feed-forward computations. By encoding token embeddings both in conventional vectors and in the amplitudes of small variational quantum circuits, our model replaces the conventional softmax attention mechanism with an entanglement-driven alternative. When trained on the Wikipedia-2 corpus using PennyLane's lightning.qubit simulator on NVIDIA A100 GPUs, QuantumGPTMini matched the perplexity of a GPT-2 Small baseline while demonstrating superior pattern recognition and symbolic reasoning behaviors. We have made this system publicly accessible through a FastAPI endpoint and a Next.js web interface. While our initial results underscore the potential of quantum-augmented language models, they also highlight challenges in moving beyond simulation, including circuit depth limitations, decoherence, and the high overhead of parameter-shift gradient estimation.

•Apr 28, 2025•0 citations

Hybrid Quantum-Classical Model for Image Classification

Muhammad Adnan Shahzad

This study presents a systematic comparison between hybrid quantum-classical neural networks and purely classical models across three benchmark datasets (MNIST, CIFAR100, and STL10) to evaluate their performance, efficiency, and robustness. The hybrid models integrate parameterized quantum circuits with classical deep learning architectures, while the classical counterparts use conventional convolutional neural networks (CNNs). Experiments were conducted over 50 training epochs for each dataset, with evaluations on validation accuracy, test accuracy, training time, computational resource usage, and adversarial robustness (tested with $ε=0.1$ perturbations).Key findings demonstrate that hybrid models consistently outperform classical models in final accuracy, achieving {99.38\% (MNIST), 41.69\% (CIFAR100), and 74.05\% (STL10) validation accuracy, compared to classical benchmarks of 98.21\%, 32.25\%, and 63.76\%, respectively. Notably, the hybrid advantage scales with dataset complexity, showing the most significant gains on CIFAR100 (+9.44\%) and STL10 (+10.29\%). Hybrid models also train 5--12$\times$ faster (e.g., 21.23s vs. 108.44s per epoch on MNIST) and use 6--32\% fewer parameters} while maintaining superior generalization to unseen test data.Adversarial robustness tests reveal that hybrid models are significantly more resilient on simpler datasets (e.g., 45.27\% robust accuracy on MNIST vs. 10.80\% for classical) but show comparable fragility on complex datasets like CIFAR100 ($\sim$1\% robustness for both). Resource efficiency analyses indicate that hybrid models consume less memory (4--5GB vs. 5--6GB for classical) and lower CPU utilization (9.5\% vs. 23.2\% on average).These results suggest that hybrid quantum-classical architectures offer compelling advantages in accuracy, training efficiency, and parameter scalability, particularly for complex vision tasks.

arXiv:2509.13353•Sep 14, 2025•1 citation

Quantum Vision Transformers

El Amine Cherrat, Iordanis Kerenidis, Natansh Mathur +3 more

In this work, quantum transformers are designed and analysed in detail by extending the state-of-the-art classical transformer neural network architectures known to be very performant in natural language processing and image analysis. Building upon the previous work, which uses parametrised quantum circuits for data loading and orthogonal neural layers, we introduce three types of quantum transformers for training and inference, including a quantum transformer based on compound matrices, which guarantees a theoretical advantage of the quantum attention mechanism compared to their classical counterpart both in terms of asymptotic run time and the number of model parameters. These quantum architectures can be built using shallow quantum circuits and produce qualitatively different classification models. The three proposed quantum attention layers vary on the spectrum between closely following the classical transformers and exhibiting more quantum characteristics. As building blocks of the quantum transformer, we propose a novel method for loading a matrix as quantum states as well as two new trainable quantum orthogonal layers adaptable to different levels of connectivity and quality of quantum computers. We performed extensive simulations of the quantum transformers on standard medical image datasets that showed competitively, and at times better performance compared to the classical benchmarks, including the best-in-class classical vision transformers. The quantum transformers we trained on these small-scale datasets require fewer parameters compared to standard classical benchmarks. Finally, we implemented our quantum transformers on superconducting quantum computers and obtained encouraging results for up to six qubit experiments.

arXiv:2209.08167•Quantum•Feb 22, 2024•162 citations
As quantum computing moves from the noisy intermediate-scale quantum (NISQ) era toward early fault-tolerant quantum computing (EFTQC), quantum-enhanced AI is shifting from a conceptual idea to experimental research. Quantum components are expected to extend the capabilities of existing AI systems, yet it remains unclear how such components will help within large multimodal models. In this study, we try to answer that question for vision-language models (VLMs) using classically simulated quantum layers. With LLaVA-OneVision as the backbone, we evaluate 4- and 8-qubit basic-ring and strongly entangling circuits at four insertion and four replacement locations. Insertion adds quantum layers to the classical model, while replacement substitutes them for classical components. Using the same setup, we compare performance and computational cost between classical and quantum-classical hybrid models. On the Visual Spatial Reasoning dataset, inserting a quantum layer after the visual projector reaches 80.7% development accuracy, 2.28 percentage-points above classical full fine-tuning. Replacing the final decoder feed-forward sublayer retains 79.22% accuracy with 13.07 million fewer trainable parameters. Quantum insertion can thus improve accuracy, and selective replacement can reduce parameters. These benefits depend strongly on where the quantum layer is inserted or which classical component it replaces: earlier replacements lower accuracy, and fewer parameters do not reliably shorten training or reduce computation. To assess potential security benefits, we evaluated the robustness of quantum-classical hybrid models under six attack settings. Overall, our study maps where quantum layers help and hurt within VLMs, offering a starting point for hybrid VLM design; whether these gains generalize remains open. The code is available at https://github.com/TITANS-Lab-Artifacts/QuantumVLM.
The paper asks whether small quantum layers can improve vision-language model accuracy, reduce computational cost, or increase adversarial robustness. It evaluates noiseless, classically simulated 4- and 8-qubit variational circuits in LLaVA-OneVision-Qwen2-0.5B, comparing two circuit families at four insertion and four component-replacement locations. Experiments use the Visual Spatial Reasoning dataset, with 7,680 training and 1,097 development examples, alongside classical fine-tuning and parameter-efficient baselines.
Under full-backbone training, inserting an 8-qubit strongly entangling circuit after the visual projector achieves 80.7% development accuracy, 2.28 percentage points above classical full fine-tuning; classical LoRA reaches 79.0%. Replacing the final decoder feed-forward sublayer achieves 79.22% while removing 13.07 million trainable parameters from approximately 893.68 million. Replacing the projector or an early decoder feed-forward sublayer instead yields approximately 51–55% accuracy. Parameter reductions do not consistently reduce measured training time or memory, and unequal early-stopping budgets prevent speedup conclusions.
Six image-perturbation attacks are evaluated on 40 development examples at ℓ∞\ell_\inftyℓ∞​ radius 4/2554/2554/255. Both iterative white-box PGD attacks reduce robust accuracy to zero for every model; no consistent robustness benefit is observed.
The stated limitations include one backbone, dataset, and training seed; development-set reuse for checkpoint selection and reporting; selection among many configurations; absent parameter-matched classical controls; and unequal training budgets. Robustness interpretation is further limited by the small subset, HSJ initialization failures, and unresolved gradient diagnostics. No quantum-hardware performance or quantum advantage is established.
Table1_Performance.csv This file contains the classification performance of the Classical CNN and the proposed SH-QC-CNN across three benchmark datasets: ImageNet-1K, CIFAR-100, and Medical X-ray. It reports Top-1 accuracy for both models and the percentage improvement achieved by the hybrid architecture. The data supports the claim that quantum-enhanced feature processing consistently improves recognition accuracy across diverse domains. Fig7_Parameter_Counts.csv This dataset reports the number of trainable parameters (in millions) for Classical CNNs and SH-QC-CNN on three datasets. It demonstrates the substantial reduction in model size achieved by replacing dense classical layers with Parameterized Quantum Circuits (PQCs), highlighting the parameter efficiency of the proposed hybrid architecture. Fig8_Training_Time.csv This file provides the total training time (in hours) for Classical CNNs and SH-QC-CNN across ImageNet-1K, CIFAR-100, and Medical X-ray datasets. The data shows that SH-QC-CNN reduces training time by approximately 34% on average, validating the computational efficiency of quantum-enhanced feature compression. Fig9_PQC_Depth_vs_Accuracy.csv This dataset captures the relationship between PQC depth (number of quantum layers) and classification accuracy. It illustrates that performance improves with depth up to three layers, after which accuracy degrades due to NISQ-era limitations such as noise and barren plateaus. The data justifies the selection of a three-layer PQC as the optimal configuration. Fig11_Noise_Robustness.csv This file contains clean and noisy accuracy values for both Classical CNN and SH-QC-CNN under Gaussian noise perturbation. It quantifies the accuracy drop and demonstrates that SH-QC-CNN is significantly more robust to input noise, confirming the stabilizing effect of quantum feature embeddings. Fig12_Qubit_Scaling.csv This dataset shows the relationship between classical feature dimensionality and the number of qubits required for quantum encoding. It evidences a logarithmic scaling behavior, indicating that high-dimensional feature spaces can be efficiently represented using a small number of qubits, making the architecture feasible for NISQ-era hardware.
Given the recent prominence of Noisy Intermediate Scale Quantum (NISQ) devices, hybrid quantum-classical neural networks have recently drawn a lot of excitement as a potential solution to improving learning efficiency. Our current paper describes a Hybrid Quantum-Classical Convolutional Neural Network (HQCNN) which is defined as a pretrained backbone of ResNet-18 and a small quantum-enhanced feature-transformation layer of Qiskit. The convolutional high-dimensional features are reduced to the classical eight-dimensional one, which in turn is encoded into a four-qubit parameterized quantum circuit. The resultant quantum expectation values are combined with the compressed features to come up with a hybridized representation of twelve dimensions, used to classify multi-class dataset. Experimental analysis of CIFAR10 data has shown that the suggested HQCNN achieves a test accuracy of 94.74% and improves class-wise precision, recall, and F1-scores than a number of existing hybrid models while using a significantly smaller number of latent features. More investigation into the quantum module indicates that it is predictable about scalability to batch, resistant to perturbations in features, and efficiently performed by evaluating circuit vectors. These results indicate that compact and expressive representations of images that can be helpful in the NISQ era can be acquired using hybrid architectures, thus providing a practical way to forward the benefits of quantum to modern deep-learning pipelines.
Code assistants for quantum computing operate exclusively on text, disregarding visual representations (circuit diagrams, Bloch spheres, and histograms) that permeate the domain’s documentation. This work is presented as a resource-and-pilot study with three contributions: (1) the first public multimodal dataset for Qiskit-oriented quantum programming, containing 8,366 samples (45% with images) across function completion, code generation, and question answering tasks, released under Apache 2.0; (2) a reproducible synthetic generation pipeline that extracts content from public Qiskit documentation, transcribes images via a multimodal model, generates question-answer pairs, and validates code through automatically generated unit tests; and (3) a feasibility study of parameter-efficient VLM specialization, in which rsLoRA (rank 32, two epochs) on Qwen3-VL-8B-Instruct is retained as the best configuration under a controlled method-selection comparison of five LoRA variants. On the text-only external benchmarks Qiskit HumanEval and Qiskit HumanEval Hard, which measure quantum-code specialization rather than visual understanding, the fine-tuned model reaches 41.72% and 30.46% Pass@1 respectively, above the IBM-specialized Qwen2.5-Coder-14B-Qiskit (40.40% / 25.83%) and granite-3.3-8b-qiskit (17.88% / 9.93%), and + 17.22 / + 15.89 pp over the Qwen3-VL-8B-Instruct base. Visual capability is assessed on the synthetic multimodal test split, where multimodal samples reach 64.29% Pass@1 versus 46.33% on text-only samples ( + 17.96 pp); a matched training ablation that withholds the multimodal portion of the training set localizes + 6.25 pp of that gap to the multimodal training signal. The dataset, models, and code are available under the Apache 2.0 license.
Transformer architectures have driven advances in natural language processing, powering systems from question answering to creative text generation. Yet as model sizes have grown from hundreds of millions to hundreds of billions of parameters, the computational cost of these systems has become prohibitive, placing advanced language technology out of reach for many researchers. In this work, we introduce QuantumGPTMini, a hybrid quantum-classical transformer with just over one million classical parameters that leverages lightweight quantum circuits to perform its core attention and feed-forward computations. By encoding token embeddings both in conventional vectors and in the amplitudes of small variational quantum circuits, our model replaces the conventional softmax attention mechanism with an entanglement-driven alternative. When trained on the Wikipedia-2 corpus using PennyLane's lightning.qubit simulator on NVIDIA A100 GPUs, QuantumGPTMini matched the perplexity of a GPT-2 Small baseline while demonstrating superior pattern recognition and symbolic reasoning behaviors. We have made this system publicly accessible through a FastAPI endpoint and a Next.js web interface. While our initial results underscore the potential of quantum-augmented language models, they also highlight challenges in moving beyond simulation, including circuit depth limitations, decoherence, and the high overhead of parameter-shift gradient estimation.
This study presents a systematic comparison between hybrid quantum-classical neural networks and purely classical models across three benchmark datasets (MNIST, CIFAR100, and STL10) to evaluate their performance, efficiency, and robustness. The hybrid models integrate parameterized quantum circuits with classical deep learning architectures, while the classical counterparts use conventional convolutional neural networks (CNNs). Experiments were conducted over 50 training epochs for each dataset, with evaluations on validation accuracy, test accuracy, training time, computational resource usage, and adversarial robustness (tested with ε=0.1ε=0.1ε=0.1 perturbations).Key findings demonstrate that hybrid models consistently outperform classical models in final accuracy, achieving {99.38\% (MNIST), 41.69\% (CIFAR100), and 74.05\% (STL10) validation accuracy, compared to classical benchmarks of 98.21\%, 32.25\%, and 63.76\%, respectively. Notably, the hybrid advantage scales with dataset complexity, showing the most significant gains on CIFAR100 (+9.44\%) and STL10 (+10.29\%). Hybrid models also train 5--12×\times× faster (e.g., 21.23s vs. 108.44s per epoch on MNIST) and use 6--32\% fewer parameters} while maintaining superior generalization to unseen test data.Adversarial robustness tests reveal that hybrid models are significantly more resilient on simpler datasets (e.g., 45.27\% robust accuracy on MNIST vs. 10.80\% for classical) but show comparable fragility on complex datasets like CIFAR100 ($\sim$1\% robustness for both). Resource efficiency analyses indicate that hybrid models consume less memory (4--5GB vs. 5--6GB for classical) and lower CPU utilization (9.5\% vs. 23.2\% on average).These results suggest that hybrid quantum-classical architectures offer compelling advantages in accuracy, training efficiency, and parameter scalability, particularly for complex vision tasks.
In this work, quantum transformers are designed and analysed in detail by extending the state-of-the-art classical transformer neural network architectures known to be very performant in natural language processing and image analysis. Building upon the previous work, which uses parametrised quantum circuits for data loading and orthogonal neural layers, we introduce three types of quantum transformers for training and inference, including a quantum transformer based on compound matrices, which guarantees a theoretical advantage of the quantum attention mechanism compared to their classical counterpart both in terms of asymptotic run time and the number of model parameters. These quantum architectures can be built using shallow quantum circuits and produce qualitatively different classification models. The three proposed quantum attention layers vary on the spectrum between closely following the classical transformers and exhibiting more quantum characteristics. As building blocks of the quantum transformer, we propose a novel method for loading a matrix as quantum states as well as two new trainable quantum orthogonal layers adaptable to different levels of connectivity and quality of quantum computers. We performed extensive simulations of the quantum transformers on standard medical image datasets that showed competitively, and at times better performance compared to the classical benchmarks, including the best-in-class classical vision transformers. The quantum transformers we trained on these small-scale datasets require fewer parameters compared to standard classical benchmarks. Finally, we implemented our quantum transformers on superconducting quantum computers and obtained encouraging results for up to six qubit experiments.