Linum.ai released Apache-2.0 weights for Pyramid-JiT (P-JiT), a pixel-space text-to-image diffusion model trained for ~660M samples (~$21K compute) as a proof-of-concept toward efficient video models. The model achieves aggressive 32×32 token reduction (4× smaller attention window), uses readout heads, PixelREPA alignment to DINOv3-L, a frozen Qwen3.5-4B text encoder, and a curriculum noise schedule to speed convergence and preserve fine detail.

The author urges readers to download and archive favorite YouTube videos because the corporate internet is unreliable, and walks through a personal, production-grade homelab setup using yt-dlp (via ytdl-sub), Jellyfin for library organization, and Tdarr for HEVC transcoding to save space. They note the tools are simple for most users but describe the quirks and complexity of integrating playlists, metadata, and re-encoding at scale.