Text-to-Image Model Trained for $21K, Apache 2.0 Weights Released
Original source
Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights)
Hacker News →A two-person research lab released a text-to-image diffusion model trained for approximately $21,000 with weights available under Apache 2.0 license. The Pyramid-JiT model represents an intermediate step toward their broader goal of democratizing video generation tools. By eliminating the separate compression model (VAE) and operating directly in pixel space, the architecture achieves significant efficiency gains: it converges using 11 times fewer training samples than the prior version while representing images with 4 times greater token efficiency.
The efficiency improvements stem from several architectural innovations. A tailored noise schedule helps pixel-space models learn fine detail more effectively. Readout heads at intermediate points along the model trunk predict lower-resolution versions of the output, strengthening gradient flow in early layers and accelerating training. The team employed curriculum learning, initially concentrating on high-noise samples to establish basic structure before expanding the noise distribution for detail refinement. Post-training optimization using a Finite Difference Flow variant further improved quality.
The model operates within clear constraints. It cannot generate text or typographic content due to deliberate filtering of text-heavy training data to focus on animation applications. Its modest 2-billion-parameter size also limits performance on complex scenarios involving crowds, multiple interactions, or extensive world knowledge—though larger models and additional post-training could address these gaps. Nonetheless, the release demonstrates that competitive image generation models can achieve reasonable quality substantially below the training budgets of typical commercial systems.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.