t2i300m
a compact text-to-image generation model
2 min readupdated
t2i300m is a 311-million-parameter Diffusion Transformer (DiT) trained from scratch by researcher SmallAICreator. It generates 512×512 images from text descriptions using a frozen flan-t5-base text encoder and a DC-AE autoencoder. Thanks to its small footprint and streamlined PixArt-style architecture, the model runs directly on consumer GPUs and can be launched in Google Colab. It is a great fit for developers, hobbyists, and anyone looking to explore how diffusion models work without needing heavy server hardware.
Key features
- Compact size: only 311M parameters (hidden size 1024, 18 layers, 16 attention heads).
- Trained from scratch: trained on a dataset of 2 million images using free Kaggle TPU sessions.
- Style variety: handles photorealism, landscapes, fantasy creatures, and artistic styles like watercolor and pixel art.
- Fast inference: generates quality images in around 30 steps using the Euler sampler.
What you can do
- Generate photorealistic scenes of food, animals, or landscapes via a Python script or the command line.
- Launch an interactive browser-based web UI directly in Google Colab.
- Create digital art and illustrations in oil painting, watercolor, or pixel art styles.
- Integrate the model and its JAX inference code into your own content generation projects.
Example
Example of generating an image with Python:
from t2i300m import T2I300M
pipe = T2I300M()
images = pipe("a red fox standing in fresh snow at sunrise", n=4, seed=2026)
images[0].save("fox.png")
Read more
DetailsArchitecture: DiT (adaLN-single, 2D RoPE, QK-norm)
Latent space: DC-AE f32c32 (frozen)
Training: 2 stages on Kaggle TPU v5e-8 (256px and 512px)
License: Apache-2.0










Comments