Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights) (www.linum.ai)

A newly released generative model, Pyramid-JiT (P-JiT), was trained for approximately $21,000, showcasing impressive text-to-image capabilities. Developed by a two-person research lab aiming to democratize animation tools, the P-JiT model serves as a proof-of-concept on the path toward more effective generative video models that address two major industry challenges: high costs and sluggish performance. By utilizing a unique pixel-space architecture that achieves significant token compression, P-JiT reduces training and inference costs, thereby making these advanced tools more accessible for creative applications. Technically, the P-JiT model differentiates itself from conventional approaches that often rely on Variational Autoencoders (VAEs) for compression. Instead, it employs diffusion models for compressing and generating images simultaneously, effectively optimizing the latent space for its tasks. The architecture converges with 11 times fewer samples compared to its predecessor while achieving aggressive token reduction. Innovations include a refined noise schedule and improved readouts along the network, which enhance gradient signals and expedite training. Although currently limited in certain areas like multi-entity interactions, the model's design suggests that scaling up and further training could overcome these hurdles, paving the way for more powerful generative tools in the near future.