Evaluating Deep Convolutional Generative Adversarial Network-Based Bird Image Synthesis at Multiple Resolutions: A Quantitative and Stability Analysis

The fast progress of artificial intelligence and deep learning has dramatically changed image generation, where Generative Adversarial Networks (GANs) represent the new standard for generating realistic images from learned data distributions. GANs employ adversarial training between a generator and a discriminator [1], which enables the model to learn to approximate complex image manifolds and generate high-quality synthetic samples. This paradigm has been used successfully in several domains, such as data augmentation, creative content creation and scientific imaging.

Among the GAN variants, Deep Convolutional GANs (DCGANs), introduced by Radford et al. [2], defined an architectural practice, including stacked convolutional layers with batch normalization and activation nonlinearities, which showed a remarkable improvement in both training stability and feature learning under unsupervised conditions. Subsequent work extended GANs to conditional scenarios, where auxiliary information such as class labels or natural language descriptions guides the generation process, enabling labeled image synthesis and text-to-image generation. In parallel, GAN-based techniques have started to influence ecological and animal-focused research, where synthetic images can help address data scarcity, class imbalance, and fine-grained classification challenges.

In this study, the investigation was conducted to use DCGAN-based architectures for synthetic bird image generation using the Caltech-UCSD Birds-200-2011 (CUB-200-2011) dataset, a standard benchmark for fine-grained avian recognition. Two models have been implemented and compared: an unconditional DCGAN that generates 64 × 64 images from random noise, and a conditional 256 × 256 DCGAN that incorporates pre-trained caption embeddings into the generation process. Our objectives are to (i) assess how image resolution and conditioning affect realism and diversity, (ii) analyze training stability across resolutions, and (iii) evaluate the models using quantitative metrics such as Fréchet Inception Distance (FID) and Inception Score (IS).

The main contributions of this work are i) making a multi-resolution evaluation of DCGAN-based bird image synthesis, comparing 64 × 64 and 256 × 256 outputs using FID and IS; ii) conducting an empirical investigation of training dynamics with a demonstration on how resolution and conditional inputs influence loss stability and visual artifacts; and iii) proposing Domain-Specific GAN (DSGAN) fine-tuning for a domain-specific application of DCGANs on fine-grained avian imagery on CUB-200-2011, with potential ecological data augmentation and wildlife image analysis benefits.

3.1 Study overview

GANs have demonstrated strong capabilities in synthesizing realistic images across various domains, including fine-grained natural categories such as birds [1, 2]. However, the performance of classical GAN architectures, particularly DCGAN, under different resolutions and conditioning strategies remains insufficiently explored in avian image synthesis. In this work, this gap has been addressed by providing a thorough analysis of DCGAN-based models trained on the CUB-200-2011 dataset [15, 16].

The research studies two dual tasks of generation:

Unconditional DCGAN (64 × 64): A baseline model that produces low-resolution bird samples from random noise itself with the architectural guidelines of a previous study [2].

Conditional DCGAN (256 × 256): A high-resolution model with pre-trained caption embeddings of the studies [1, 3] and multi-stage refinement like techniques of the study [5].

Both architectures are trained adversarially with the binary cross-entropy loss and measured using FID for samples generated from the generator, and IS, a standard in assessing generative quality [4]. The paper also investigates the stability of training, loss dynamics and qualitative differences between generated and real images.

This unified pipeline of dataset preparation, model training and evaluation offers a well-organized form that facilitates understanding of the role played by image resolution and conditioning on DCGAN-based systems as related to fine-grained avian image generation.

3.2 Research objectives

The main goals for this study are as follows: (a) to assess the performance of DCGAN-like architectures in generating synthetic bird images in multiple resolutions. This is realized by addressing the following objectives of the study:

• Dataset preparation: Preprocess the CUB-200-2011 dataset [15, 16], normalizing images, resizing them and integrating pre-trained caption embeddings for conditional generation.
• Model development: Apply two variants of DCGANs, unconditional (64 × 64) and conditional (256 × 256), according to well-established modelling criteria [2, 3].
• Training and stability analysis: Train both models with the same hyperparameters, analyze stability for training dynamics such as the loss of the generator and the discriminator to compare resolution.
• Quantitative evaluation: Quantitatively evaluate generative performance via FID and IS as with best practices in GAN evaluation [4] across model variants.
• Qualitative assessment: Visually inspect the sampled images to evaluate how realistic, diverse, and fine-grained (aligned with text description) the generated samples are, following qualitative evaluations from previous text-to-image [1] and multi-stage GAN studies [5].
• Domain-specific insight: Give some concrete examples on the practicality of utilizing DCGANs for applications in ecology and biodiversity by recent work in avian classification and augmentation [6, 7].

These goals altogether make a complete analysis of DCGAN performance in fine-grained bird image generation and set a benchmark for further applications with newer generative models.

This section details the implementation of the generative pipeline to ensure reproducibility. The overall workflow, illustrated in Figure 2, encompasses sequential stages from the dataset loading and preprocessing to model evaluation.

2.png

Figure 2. Implementation workflow

This workflow is consistent with established GAN training practices [2, 4] and text-to-image conditioning frameworks [1, 3, 5].

5.1 Model architectures

Two DCGAN variants were implemented: an unconditional 64 × 64 model and a conditional 256 × 256 model. Both follow the architectural guidelines introduced by Radford et al. [2], with modifications to support conditioning.

5.1.1 Unconditional DCGAN (64 × 64)

Generator architecture: The generator maps a 100-dimensional noise vector z ~ N (0,1) to a 64 × 64 × 3 RGB image (Table 2).

This architecture is identical to the canonical DCGAN generator [2].

Discriminator architecture: The discriminator maps a 64 × 64 × 3 image to a scalar probability (Table 3).

This structure follows the DCGAN discriminator design [2].

Table 2. Generator architecture (64 × 64 × 3)

Table 3. Discriminator architecture (64 × 64 × 3)

5.1.2 Conditional DCGAN (256 × 256)

The conditional model incorporates caption embeddings into both generator and discriminator (Tables 4 and 5), inspired by CGAN [3] and StackGAN [5].

Table 4. Generator architecture (256 × 256 × 3)

Table 5. Discriminator architecture (64 × 64 × 3)

Caption embedding integration:

• Each image has a 1024-dimensional text embedding (from the dataset’s .pkl files).
• Embeddings are normalized and linearly projected to 128 dimensions.
• The embedding is concatenated with the noise vector:

z' = [z; c]

where, $z \in \mathrm{R}^{100}, c \in \mathrm{R}^{128}, z^{\prime} \in \mathrm{R}^{228}$.

This deep upsampling stack enables high-resolution synthesis, similar to Stage‑II StackGAN [5].

Discriminator architecture (256 × 256): The discriminator jointly processes the image and caption embedding.

This joint embedding approach follows CGAN principles [3].

5.2 Training strategy and stability techniques

Both models were trained using the same optimization strategy to ensure comparability (Table 6).

Table 6. Hyperparameters

These hyperparameter settings follow best practices for stable GAN training [2, 4]. The training loop for each batch proceeded as follows:

(1) Discriminator update

Real images x were sampled from the dataset, and fake images G(z) were generated from random noise z. The discriminator loss was computed as:

L_p = -E[logD(x)] – E[log(1 – D(G(z)))]

The discriminator weights were then updated via backpropagation.

(2) Generator update

Fake images were generated, and the generator loss was computed as:

L_G = -E[logD(G(z))]

The generator weights were subsequently updated.

(3) Conditional training (56 × 256 model)

The training procedure followed the CGAN paradigm [3]. Here, the embeddings are fed into both networks (discriminator and generator). The discriminator received image-embedding pairs as input, while the generator learned to align visual features with semantic cues.

Since GANs are unstable, several stabilization strategies were employed, consistent with the recommendations in GAN literature [4]:

• Batch normalization: Applied to every layer of both the generator and discriminator, with one exception.
• Label smoothing: Substitute actual labels for 0.9 instead of 1.0; makes the discriminator bluff more difficult when there are ample examples.
• Gradient monitoring: Report explosive gradients were discovered early and abated by lowering the learning rate.
• Periodic sample generation: Keep an image for every epoch, so that a collapse in mode can be detected.
• Embedding normalization: Applied to caption embeddings to taper gradient spikes.

5.3 Evaluation metrics

Two widely accepted metrics were used: FID, which measures distribution similarity between real and generated images, and IS, which measures image quality and diversity. Both metrics are standard in GAN benchmarking [4].

Table 7. Tools and libraries

These tools align with common deep learning workflows (Table 7).

The unconditional 64 × 64 DCGAN and the conditional 256 × 256 DCGAN were quantitatively benchmarked, analyzed for training stability, visually inspected and compared with other contemporary generative baselines. The results are presented thematically to focus the reader's attention on the role of resolution and conditioning in fine-grained avian image generation.

7.1 Quantitative evaluation

The generative quality of the samples is evaluated by FID score and IS score (Table 8). Our conditional 256 × 256 model showed an FID of 52 and an IS of 4.1, surpassing the unconditional 64 × 64 one (FID 65, IS 3.4). The lower FID shows that the conditional model generated feature distributions were closer to real bird images, and the higher IS indicates diverse and confident classification of generated samples.

Table 8. Quantitative results

7.2 Training loss behavior

Stability of training was examined on the basis of loss curves for the discriminator and the generator. The 64 × 64 failed to achieve stable loss (0-10) in discriminator terms, signifying adversarial imbalance, i.e., an occasional mode collapse. The generator loss had distinct spikes, suggesting the difficulty in learning subtle features. By contrast, the 256 × 256 conditional model achieved more steady convergence: the discriminator loss levels off at about 4–6 after ~30 epochs, while the generator continues to fall slowly enough to allow for coherent textures and species-specific characteristics. Conditioning helped the gradient flow and alleviated instability, demonstrating the benefits of semantic guidance.

7.3 Qualitative evaluation

Subjective evaluation supported the result quantitatively. The 64 × 64 model captured primary silhouettes and coarse color distributions with blurry and anatomically distorted polishing and background artifacts. The edge of conditional 256 × 256 model was relatively sharp and the texture details, such as feathers, were clear; species-level features (wing stripe, beak shape) were more realistic. Some minor artifacts, including color bleeding and checkerboard effects, were visible; nevertheless, the quality of the enhanced images was significantly superior. Visual comparisons showed that the conditional outputs were visually closer to real CUB-200-2011 images, and some generated images resembled low-quality real photos.

7.4 Diversity analysis

The generalizability was measured with random batch inspection, IS variance and latent space interpolation. The 64 × 64 one generated repetitive patterns and small color palettes, rather than a full-color animal, whereas the 256 × 256 one produced diverse poses, species-specific coloration and more complicated backgrounds. Conditioning facilitated the exploration of a wider manifold of bird-like appearances, increasing ecological validity.

7.5 Comparison with modern baselines

Although DCGAN is an early generative architecture, it remains a practical and computationally efficient baseline, being effective and easy to interpret. We compare our results in Table 9 to state-of-the-art generative models. Although StackGAN has FID < 35 by multi-stage refinement, StyleGAN2 can reach FID < 10 with state-of-the-art realism and in-diffusion models are the current leaders with FID of < 5. This comparison places DCGAN as an applicable baseline rather than a state-of-the-art model, which is particularly useful in ecological augmentation experiments where speed of computation is important.

Table 9. Comparison with modern generative models

7.6 Summary of key findings

• The conditional 256 × 256 DCGAN significantly outperformed the unconditional 64 × 64 model on FID and IS.
• High-resolution and semantically conditioned inputs led to a more stable training process and balanced adversarial learning.
• The 256 × 256 model generated images with sharper textures, clearer structural details, and greater visual diversity.
• Although DCGAN does not achieve the performance of state-of-the-art architectures, it remains an efficient and interpretable baseline for ecological image synthesis tasks.
• This study fills a gap in the literature by presenting a multi-resolution, stability-oriented assessment of DCGANs using fine-grained avian imagery.

7.7 Results visualization

To complement the quantitative and qualitative analyses, a series of grouped visualizations was prepared to illustrate the training process and outcomes of both DCGAN variants.

Examples from the CUB-200-2011 dataset at resolutions of 64 × 64 and 256 × 256 and samples drawn from the unconditional and conditional DCGAN model are shown in Figures 3-6. These panels contrast the resolution and conditioning of where we point out that the conditioned 256 × 256 model produces a better texture (sharper) and species-level feature (more coherent) than the blurred silhouettes formed during 64 × 64 models.

Figures 7 and 8 show the generator and discriminator loss curves of the two models. The 64 × 64 DCGAN was unconditionally trained and often suffered from oscillations and instability, whereas the 256 × 256 conditional case showed more stable convergence with fewer fairness issues. These plots support the stability enhancements reported in Section 8.2 visually.

Examples are compared side by side with real bird images and synthetic counterparts in Figures 9 and 10. The generative capacity of the model is demonstrated by samples it generates in unconditional mode, which are small thumbnail-like blurry blobs of color with little discernible content compared to naturalistic images, and in conditional mode, images that exhibit recognizable properties like wing-stripes, beak shapes and feather textures. These comparisons and higher conditioning bring the output energy spectrum close to population data.

This narrative transcends through these visualizations: data preparation, two epochs of the training process and after convergence. They are qualitative evidence that reinforces the quantitative gain in FID and IS, and they show the practical usefulness of conditional DCGANs for fine-grained bird image generation applications.

Figure 3. Training data (64 × 64 × 3)

Figure 4. Training data (256 × 256 × 3)

Figure 5. A batch of generated sample data (64 × 64 × 3)

Figure 6. A batch of generated sample data (256 × 256 × 3)

Figure 7. Generator and discriminator loss during training (64 × 64 × 3)

Figure 8. Generator and discriminator loss during training (256 × 256 × 3)

Figure 9. Real images vs. fake images comparison (64 × 64 × 3)

Figure 10. Real images vs. fake images comparison (256 × 256 × 3)

7.8 Challenges

Several technical challenges emerged during the study:

a) Missing text descriptions

The dataset provided caption embeddings but did not include the original textual descriptions due to privacy restrictions. This limitation restricted the ability to validate embedding semantics, a constraint also noted in prior text-to-image work [1].

b) Computational constraints

Training 256 × 256 GANs is computationally demanding. Limited GPU resources restricted the scope of hyperparameter exploration and prolonged debugging cycles, consistent with challenges reported in GAN training literature [4].

c) Sensitivity to hyperparameters

DCGANs are highly sensitive to hyperparameter configurations, including learning rate, batch size, and normalization strategies. In our early experiments, getting these wrong often led to common problems like discriminator overpowering, generator collapse and gradient instability. Addressing these issues required iterative tuning of architectural components and optimization settings.

d) Stability issues at low resolution

The 64 × 64 model exhibited training instabilities, confirming that low-resolution GANs struggle to pick up smaller, finer-grained features. These problems occurred mainly in datasets with high intra-class variability, such as CUB-200-2011 [15, 16]. These challenges provided valuable insights into adversarial training dynamics and informed the final design of our model.