RUPASTRA Text-to-Image 10M
an ultracompact model for pixel art generation
2 min read
A miniature AI model with just 10.23 million parameters trained from scratch to create pixel art characters from text descriptions. The model outputs images at a resolution of 128x128 pixels based on the m1guelpf/nouns dataset. Despite its modest size (~40 MB), it uses interesting architectural solutions: Grouped Query Dual-Attention (GQDA, an attention acceleration method), RoPE-2D positional embeddings, and an autoregressive approach. The development is great for researchers and hobbyists who like running lightweight generative models locally.
Key features
- Minimal weight: only ~10.23 million parameters and a weight file size of about 40 MB.
- GQDA architecture: a custom transformer with Grouped Query Dual-Attention and SwiGLU layers.
- Pixel art: specialized in generating stylized characters with a unique combination of shape and color.
- Autoregression: processing via 2x2 Spatial Patch Decoding and a compact BPE tokenizer.
What you can do
- Generate a batch of unique characters for an indie game in the Nouns style.
- Explore the behavior of miniature autoregressive image generation models.
- Run graphics generation on very weak hardware thanks to tiny resource consumption.
Example
Example text prompt for character generation:cat-shaped head, red-colored body, warm background
Read more
DetailsTraining parameters and architecture:
• Architecture: 8 layers, d_model = 288, FFN dim = 768
• Training: 12 epochs, 9,348 steps, CFG dropout 0.1
• Metrics: Train Loss 1.1018, Eval Loss 0.9337
• Tokenizer: custom BPE with 828 words










Comments