acceptodds
Under review as a conference paper at ICLR 2027

SQGen: Structured Quantum Image Generation with Latent-Modulated Quantized Tensor Trains

Abstract

Generating images directly from quantum systems is an attractive but unresolved goal on NISQ hardware. Existing quantum generators face several coupled obstacles: barren plateaus that block trainability, expensive quantum circuit preparation, and hardware noise that erodes quantum information with depth. A further difficulty is generating image-scale outputs without a classical decoder. We propose SQGen, a decoder-free quantum generator whose circuit follows a quantized tensor train (QTT) skeleton with latent modulation. Specifically, SQGen promotes the QTT bond index of the target pixel distribution to ancilla bond qubits, so that each circuit site acts on a bond register, the two physical qubits that carry the row- and column-bit of one image scale, and a coarse-to-fine link to the previous scale. The latent enters through latent modulation: an exact angle-level factorization of each re-uploading rotation into a trainable main path plus an additive latent term, reducing to the main path when the latent term is disabled. During training, we create a differentiable model in the classical system under gate-compatibility constraints, with a torus prior as the latent distribution. Every trained operator maps one-to-one to a native quantum gate, yielding a compact, deployable quantum circuit with no classical decoder in the inference path. Extensive experiments on image datasets and synthetic data demonstrate that SQGen trains stably, generates images end-to-end from a shallow circuit with no classical decoder, and shows promising feasibility on real quantum hardware.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum machine learning, Quantum image generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_xGFhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to address a very specific question: is it possible to generate images directly using a shallow quantum circuit, bypassing the need for a classical neural network decoder? The authors propose SQGen, a method that embeds the multi-scale structure of Quantized Tensor Trains (QTT) directly into the quantum circuit and employs latent modulation to control the output, while performing the training process on a classical computer.

Critical strengths

The experimental design is comprehensive and effectively validates the roles of the various design components. In addition to image generation experiments on MNIST and Fashion-MNIST, the paper incorporates synthetic targets with well-defined structures to specifically evaluate the model's ability to capture multi-scale spatial structures and latent-conditioned variations. However, the experiments did not include comparisons with recent methods; only internal baseline ablation studies were conducted. Therefore, I am not inclined to accept the paper.

Critical weaknesses
  1. The experimental comparison is not sufficiently strong.

  2. The architectural novelty relative to prior QTT-based quantum circuit designs is not sufficiently clarified. And the practical cost of decoder-free generation is insufficiently analyzed.

These obvious weaknesses incline me toward a weak rejection.

Questions for authors
  1. While a decoder-free approach eliminates the need for a CNN, it introduces costs associated with shot complexity and probability estimation; this issue becomes particularly pronounced as resolutions increase.

2.I believe it is necessary to compare the proposed method against recent quantum image generation approaches, such as QIREN [1], QINR-QGAN [2], OQIDDM [3], and RQ-QGAN [4], to more convincingly demonstrate the effectiveness and competitiveness of the proposed quantum circuit design.Comparisons with decoder-free or coordinate-conditioned image generation methods are also required. Use classical QTT + LM to isolate the respective contributions of "QTT itself" and the "quantum" aspect; Additionally, include a resource-matched generic QCBM + LM to isolate the topological inductive bias.

  1. The circuit architecture of SQGen appears closely related to the QTT-based structured quantum circuits introduced in TNQE [5]. In particular, both methods employ a multiscale QTT organization, promote the bond dimension to ancillary bond qubits, and construct local circuit blocks acting jointly on the bond register and scale-wise physical qubits. I would therefore like the authors to clarify how much of the circuit architecture is newly introduced in SQGen compared with TNQE, and which architectural components are specific to the generative setting.

[1] Jiaming Zhao, Wenbo Qiao, Peng Zhang, and Hui Gao. Quantum implicit neural representations. In Proceedings of the 41st International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, 235:60940–60956, 2024.

[2] Han Qi, Yihan Xu, Hao Wang, Abdullah Gani, and Lip Yee Por. RQ-QGAN: Recurrent quantum generative adversarial network with QLSTM-inspired hidden states. The Journal of Supercomputing, 82:475, 2026.

[3] Jiale Zhang, Xilong Che, Yuzhe Fan, Shun Peng, Geng Chen, Quangong Ma, and Juncheng Hu. Denoising diffusion models with optimized quantum implicit neural networks for image generation. Future Generation Computer Systems, 173:107875, 2025.

[4] QuanGong Ma, ChaoLong Hao, NianWen Si, Geng Chen, Jiale Zhang, and Dan Qu. Quantum adversarial generation of high-resolution images. EPJ Quantum Technology, 12:3, 2025.

[5] Guang Lin, Toshihisa Tanaka, and Qibin Zhao. Structured Unitary Tensor Network Representations for Circuit-Efficient Quantum Data Encoding. Advances in Neural Information Processing Systems (NeurIPS), 2026.

Reviewer_ouYHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a quantized tensor train-based quantum image generating technique that trains classically and can perform inference without a classical decoder in the critical path. The technique introduces a method to map a QTT bond skeleton to a quantum circuit in addition to a reuploading approach for inputting the latent onto the quantum circuit. The technique is evaluated on MNIST, Fashion-MNIST, and synthetic datasets in classical simulation where it is demonstrated to improve performance over selected baselines. Then, the technique is deployed to real quantum hardware to test its feasibility.

Critical strengths

The technique to use a quantized tensor train bond skeleton inside a quantum circuit for image generation appears novel and could be of interest to further quantum machine learning research. I see this the strongest aspect of this paper as it could provide tools for future research to expand in this domain.

Critical weaknesses

The selection of baselines is limiting. In order to demonstrate utility of this approach empirically, I would have liked to see comparisons with stronger comparable quantum baselines such as Huang et al., 2021. The lack of clarity on how this paper compares to other established qml image generation techniques impacts my decision as it is not clear this is empirically advantageous.

The paper mentions that the bulk of modeling capacity for comparable methods is in a classical decoder which is a core piece of the paper’s motivation. However, this is not accurate in the paper’s cited work of Huang et al., 2021, where no classical decoder is in the inference path. This limits a key motivating argument of this work as it does not provide a unique advantage over competing works.

The real hardware experiment adds a new baseline only for this section and does not compare against any of the previous baselines. The choice of only adding in amplitude embedding for this comparison limits the impact of the real machine experiment as it is known that amplitude embedding adds many more gates and will likely accumulate more noise, but the other baselines, which do not introduce this problem, are omitted.

Questions for authors

Why are KL divergence and L1 divergence the key evaluation metrics instead of more direct assessment of image quality like Inception Score?

How does the technique perform compared to established quantum models such as Huang et al., 2021?

Why does each core technique perform worse on its own and only better when combined with one another?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.