GANs fight back: review maps how adversarial networks still shape text-to-image synthesis

A systematic review published in Discover Artificial Intelligence on 2 September 2026 argues that generative adversarial networks, though eclipsed on public leaderboards by diffusion and autoregressive models, remain a scientifically active and practically relevant branch of text-to-image synthesis. The paper, authored by Shreedatta S. Sawant, S. Kaliraj, S. Raghavendra and Rudresh Shirwaikar, is a PRISMA-aligned survey covering 19 GAN-based text-to-image models evaluated across six datasets and six evaluation metrics, with a primary literature corpus spanning January 2019 to December 2023.

The review’s methodology was deliberately structured. The authors searched five databases — IEEE Xplore, ACM Digital Library, Google Scholar, arXiv and SpringerLink — using Boolean strings combining generative adversarial network terms with text-to-image and deep learning vocabulary. From 312 records retrieved, 187 were screened on title and abstract, 74 were assessed for full-text eligibility, and 36 studies entered the primary corpus, supplemented by pre-2019 foundational references and three emerging 2024–2025 works handled separately. Inclusion required peer-reviewed English articles directly addressing GAN-based text-to-image synthesis with quantitatively reported results.

Four main contributions anchor the paper. The authors propose a tri-modal input taxonomy unifying general text-based, dialogue-driven and scene-layout-conditioned synthesis under one framework. They assemble what they describe as the most thorough cross-metric performance comparison for GAN-based text-to-image generation to date, spanning 19 models, six metrics and six datasets. They introduce a secondary five-dimensional taxonomy characterising each model by text encoder type, cross-modal alignment mechanism, generator–discriminator design, training stabilisation strategy and controllability. Finally, they provide a formal mathematical treatment of five core GAN training objectives, from the standard minimax game through conditional GANs, Wasserstein GAN with gradient penalty, mode-seeking regularisation and the Deep Attentional Multimodal Similarity loss used in AttnGAN and its variants.

The datasets underpinning the field reveal a striking concentration. CUB-200, the Caltech-UCSD Birds collection of 11,788 images across 200 bird species, appears in 44.44% of reviewed studies, followed by the Oxford-102 Flower dataset with 8,189 images in 103 classes at 27.78%, and MS-COCO at 19.44%. The remaining benchmarks — FashionGen with 293,008 high-definition fashion images, the RSICD remote sensing captioning set with 10,921 images, dialogue-based CoDraw and handwritten-digit MNIST — each account for roughly 3% of usage. The authors flag this imbalance as a potential evaluation bias: models achieving high Inception Scores on the narrow, single-category CUB domain may not generalise to MS-COCO’s complex multi-object scenes.

The survey’s central analytical finding concerns what drives performance. Encoder capacity and cross-modal alignment strategies emerge as the primary determinants of cross-dataset generalisation, while attention mechanisms and discriminator design most strongly influence image quality and semantic consistency. In the text-utilisation category, BERT-based encoding reportedly outperforms LSTM-based approaches by more than 10 Inception Score points on MS-COCO, yet self-attentional approaches such as RifeGAN2, with an Inception Score of 5.23, lead on the fine-grained CUB benchmark — likely because the limited bird domain makes large-scale pre-training less critical. Mode-seeking regularisation, as implemented in SAM-GAN, is identified as the most reliable protection against mode collapse across dataset types.

Architectural comparisons sharpen the picture further. On the common CUB-200 benchmark, DualAttn-GAN achieves the lowest reported Fréchet Inception Distance at 14.06 through dual-axis attention, while SS-TiGAN reaches a comparable 14.20 using stacked discriminators at two resolutions — evidence, the authors argue, that scaling discriminator supervision matters more than generator complexity. In the output-control category, a clear trade-off surfaces: ControlGAN delivers the highest word-level controllability but the lowest category Inception Score at 4.58, MTC-GAN maximises diversity with multiple generators but posts the weakest FID at 45.12, and DiverGAN strikes the best balance with an Inception Score of 4.98 and an LPIPS of 0.682. Across all surveyed models, no approach simultaneously achieves high controllability and high diversity.

Multistage architectures receive dedicated analysis. Models that progressively refine images across stages — Struct-GAN scoring 5.84 on Inception Score, MA-GAN at 4.76 and SAM-GAN at 4.61 — demonstrate strengths in spatial coherence and detail refinement, but the review notes that PCCM-GAN’s pyramid structure exposes a persistent resolution–semantic consistency trade-off: as resolution climbs the pyramid, semantic consistency degrades. Self-attention mechanisms, whether along channel, pixel or word axes, consistently improve both Inception Score and FID relative to non-attentional baselines, though they raise computational demands.

Evaluation methodology itself emerges as an unsolved problem. The review catalogues nine metrics — Inception Score, FID, Visual Semantic Similarity, R-Precision, LPIPS, accuracy, Classification Accuracy Score, Rank-1 and human ranking — and concludes that no single metric suffices. IS correlates with diversity but is sensitive to class imbalance and can miss mode dropping; FID captures distributional alignment but is influenced by resolution mismatches and the specific Inception checkpoint used; R-Precision can be inflated when models overfit the retrieval embedding space. The authors recommend that future studies report at minimum IS, FID and one semantic alignment metric, complemented by human evaluation wherever feasible.

The limitations section is candid. GAN training remains notoriously unstable, with mode collapse, oscillation and non-convergence endemic to the adversarial setup. High-resolution generation above 256×256 pixels is virtually absent from the GAN text-to-image literature, leaving open whether design conclusions drawn at low resolution transfer upward. Cross-domain transfer without fine-tuning was assessed in only 2 of the 19 surveyed models, and most original papers omit GPU specifications, limiting reproducibility. Ethical concerns around deepfakes and misuse are also noted, alongside the need for detection tools and policy regulation.

Context within the broader generative landscape frames the review’s conclusion. Since 2022, diffusion models such as DALL-E 2, Stable Diffusion and Imagen, along with autoregressive transformers like PARTI, have reported state-of-the-art results on major text-to-image benchmarks, frequently outperforming GANs on FID and human preference evaluations. Yet the authors document continued GAN activity in 2024–2025, including ML-GAN’s multi-level text conditioning, a Sparse Mixture-of-Experts GAN that dynamically routes specialised sub-networks, and multimodal frameworks combining text, image and style signals. GAN-based methods retain decisive advantages in inference speed, fine-grained controllability and resource efficiency, the review concludes, keeping them relevant for production and edge deployment — while six open directions, from cross-domain generalisation and unified controllability metrics to hybrid GAN–diffusion architectures and fairness evaluation, define the field’s remaining frontier.

Subject of Research: A systematic review of generative adversarial network architectures for text-to-image synthesis

Article Title: Exploring the evolution of generative adversarial network architectures for text to image synthesis a comprehensive review

Article References: Sawant, S. S., Kaliraj, S., Raghavendra, S., & Shirwaikar, R. (2026). Exploring the evolution of generative adversarial network architectures for text to image synthesis a comprehensive review. Discover Artificial Intelligence, 6(1), Article 1027. https://doi.org/10.1007/s44163-026-01991-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01991-w

Keywords: generative adversarial networks, text-to-image synthesis, systematic review, deep generative models, attention mechanisms, mode collapse, evaluation metrics, diffusion models, CUB-200, MS-COCO, controllability, image generation