{"version":1,"type":"story","url":"https://digestai.news/story/two-person-lab-releases-text-to-image-model-trained-for-21k","json":"https://digestai.news/story/two-person-lab-releases-text-to-image-model-trained-for-21k.json","markdown":"https://digestai.news/story/two-person-lab-releases-text-to-image-model-trained-for-21k.md","slug":"two-person-lab-releases-text-to-image-model-trained-for-21k","headline":"Two-person lab releases text-to-image model trained for $21K","summary":"A two-person research lab shared an experimental text-to-image model trained for about $21,000. The model, built with a **Pyramid-JiT (P-JiT)** architecture, compresses image data more aggressively than typical diffusion models, reducing training costs and improving detail generation. It uses a **2B-parameter DiT** backbone with a **4B-parameter text encoder** (Qwen3.5-4B) and supports 1024×1024 resolution outputs.\n\nThe lab, which focuses on democratizing animation tools, notes the model’s limitations: it avoids text-heavy data (no rendered text, posters, or books) and struggles with complex scenes like full-body poses or crowds. Post-training with **Finite Difference Flow Optimization (FDFO)** and human feedback was applied to refine outputs. The model’s weights are released under **Apache 2.0**, and the lab plans to iterate toward video models next.","keyPoints":["Two-person lab trained a text-to-image model for ~$21K, releasing weights under Apache 2.0","Uses P-JiT architecture to compress images aggressively, improving detail and reducing training costs","Model lacks text-heavy data support and struggles with complex scenes like crowds or multi-person interactions"],"whyItMatters":"This model demonstrates how smaller teams can achieve competitive image generation at low cost, pushing the boundaries of what’s possible with minimal resources. It could inspire others to explore cost-efficient architectures for creative tools.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":["Qwen"],"models":["Linum v3","P-JiT","Qwen3.5-4B","DINOv3-L","Claude Opus 5.5"],"people":[]},"firstPublishedAt":"2026-11-04T16:45:11Z","updatedAt":"2026-11-04T16:45:11Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"linum.ai","title":"Show HN: A decent text-to-image model trained for ~$20K (Apache 2.0 Weights)","url":"https://linum.ai/field-notes/pjit-21k","publishedAt":"2026-11-04T16:45:11Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Hacker News","url":"https://news.ycombinator.com/item?id=50368208","points":42}],"thread":null,"cite":{"text":"Digest AI, \"Two-person lab releases text-to-image model trained for $21K\", 4 November 2026, https://digestai.news/story/two-person-lab-releases-text-to-image-model-trained-for-21k","publisher":"Digest AI","title":"Two-person lab releases text-to-image model trained for $21K","datePublished":"2026-11-04T16:45:11Z","url":"https://digestai.news/story/two-person-lab-releases-text-to-image-model-trained-for-21k"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}