InstantPainting: Expanding GANs for Efficient Text-Conditioned Image Generation Platform

Bing-Kun Bao, Yefei Sheng, Jie Wang, Yaning Li, Sisi YouAAAI 2025 · DOI

Presented at AAAI-25, InstantPainting introduces a text-conditioned image generation platform that challenges the current dominance of resource-heavy Diffusion Models. By pivoting back to a highly optimized Generative Adversarial Network (GAN) architecture, the researchers demonstrate that high-quality image synthesis can be achieved with significantly lower computational overhead and faster generation speeds.

The Efficiency Challenge in Text-to-Image Generation

The current state-of-the-art in text-to-image (T2I) generation is dominated by Large-scale Diffusion Models like SD-XL and DALL-E 3. While these models produce impressive results, they suffer from three primary drawbacks:

  1. Prohibitive Costs: Training and deploying these models requires massive GPU clusters, often costing millions of dollars.
  2. High Latency: The iterative sampling process inherent to diffusion makes real-time generation difficult.
  3. High Barriers to Entry: These requirements effectively lock out smaller research teams and startups from developing localized T2I services.

InstantPainting addresses these issues by proving that a streamlined GAN-based approach can provide a competitive experience while being orders of magnitude more efficient.

System Architecture and Methodology

InstantPainting is built upon an extended version of GALIP (Generative Adversarial CLIPs). The authors implemented several key architectural and procedural improvements to bridge the quality gap between GANs and Diffusion Models.

Overview of the InstantPainting generation pipeline, backend architecture, and multi-platform WebUI.
  • Architectural Extensions: The system augments the generator and discriminator with a convolutional network-based module and a specialized noise selection module to improve image stability and detail.
  • CLIP Upgrade: By swapping the standard CLIP model for a more robust ViT-16 backbone, the system achieves better alignment between complex text prompts and visual features.
  • Multi-Stage Fine-Tuning: To handle noisy web-scale data, the authors utilized a multi-stage training scheme with screening strategies to ensure the model learns from high-quality image-text pairs.
  • Lightweight Deployment: The platform uses a Flask-based backend and an SQLite database, allowing it to run efficiently on a small cluster of consumer-grade RTX A6000 GPUs rather than massive industrial clusters.

Data Strategy: Pre-training and Fine-tuning

The model achieves its efficiency by utilizing a fraction of the data typically required for T2I:

  • Pre-training: It leverages a subset of the LAION-400M dataset, using only about 3% of the data volume compared to larger competitors.
  • Fine-tuning: The authors curated a specialized dataset of approximately one million image-text pairs to refine the model for specific high-demand categories like anime, food/recipes, and landscapes.

Performance and Key Results

InstantPainting demonstrates a striking reduction in resource requirements while maintaining high performance. As shown in the comparison below, the model operates with significantly fewer parameters and lower deployment costs.

Method Parameters Datasets Deployment Cost
SD-XL 6.6B 5B >256 Tesla V100
DALL-E 3 3.5B 5B 1024 Tesla V100
InstantPainting 0.23B 400M 16 RTX A6000

Beyond cost, the platform delivers on speed and versatility:

  • Speed: Image generation is roughly 4 to 5 times faster than diffusion-based alternatives.
  • Creative Control: The system handles complex prompts and niche tasks, such as poetry-to-image synthesis and "dark cooking" (generating intentionally unappetizing food images).
  • Output Quality: Despite its small size (0.23B parameters), the model supports high-resolution outputs up to 1024x1024.
Sample outputs showing variations of an anime-style character, demonstrating the model's ability to maintain style consistency and detail.

Conclusion and Future Work

InstantPainting proves that GANs remain a powerful and relevant tool for democratizing AI deployment. By focusing on domain specialization and architectural optimization, the researchers have created a platform that wins decisively on speed and cost without sacrificing the creative potential of text-to-image synthesis.

Future work will likely focus on expanding the model's general "world knowledge" to handle more obscure prompts and conducting more rigorous human-centric evaluations to compare the subjective "beauty" of its outputs against larger diffusion-based models.

Summary updated

BibTeX
@inproceedings{bao2025instantpainting,
  author = {Bing-Kun Bao and Yefei Sheng and Jie Wang and Yaning Li and Sisi You},
  title = {{InstantPainting: Expanding GANs for Efficient Text-Conditioned Image Generation Platform}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35345},
}
Figure