{"id":15231,"date":"2026-11-17T11:03:37","date_gmt":"2026-11-17T02:03:37","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=15231"},"modified":"2026-11-17T11:03:37","modified_gmt":"2026-11-17T02:03:37","slug":"pyramid-jit-2b-text-to-image-model","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/pyramid-jit-2b-text-to-image-model\/","title":{"rendered":"Pyramid-JiT: A 2B Text-to-Image Model Trained for $21K"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Published<\/td>\n<td>2026-11-05<\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Unverified (not confirmed by a primary source)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>The information covered in this article is based on unconfirmed reports that have not been officially verified.<\/p>\n<p>It is reported that Linum AI, a two-person research lab, has announced a 2B parameter text-to-image model named Pyramid-JiT (P-JiT), which was trained at an estimated compute cost of about $21,000, and released its model weights under the Apache 2.0 license. The model is reported to adopt a novel architecture that performs direct compression and diffusion in pixel space without using a traditional Variational Autoencoder (VAE).<\/p>\n<h2>Why It Is Making Noise<\/h2>\n<p>At a time when building models by throwing massive compute resources and high budgets at them has become commonplace, it has drawn significant attention that a text-to-image model with practical quality was built on an extremely low budget of about $21,000 (8,415 H100-hours).<\/p>\n<p>In particular, strong technical interest has been sparked by reports that it applies the &#8220;Pyramid-JiT&#8221; architecture, which skips the VAE compression used as a separate model in typical image and video generation models, instead having the Diffusion Transformer (DiT) itself simultaneously perform latent space learning and generation. It is explained that this approach achieves convergence in 1\/11th the number of samples compared to standard models (such as Linum v2 using Wan VAE and DiT) and realizes 4x more aggressive token reduction (32&#215;32 token reduction). The fact that the model weights have been released as open weights on Hugging Face and elsewhere is also a factor driving discussion in the community.<\/p>\n<h2>Discussion Points<\/h2>\n<p>In community discussions, opinions are being exchanged from multifaceted perspectives, combining technical admiration for the presented results with challenges in drawing accuracy and the provenance of the training data. The main discussion points are as follows:<\/p>\n<h3>Evaluation of the Pixel-Space Architecture Eliminating VAE<\/h3>\n<p>Attention is drawn to the novelty of the approach, which eliminates the VAE that was previously common in model building and instead directly performs pixel-space compression and prediction inside the DiT. Positive reactions can be seen regarding the ingenuity that achieved this level of generation capability on a low budget.<\/p>\n<h3>Rendering Quality and Specific Challenges<\/h3>\n<p>Based on published sample images and failure cases, points are being raised about weaknesses in rendering. Discussions note that typical problems inherent in existing text-to-image models remain in this model as well, such as unnatural hand rendering, broken limbs, excessive image smoothing (over-smoothing), and artificial glossiness.<\/p>\n<h3>Origin and Content of the Training Dataset<\/h3>\n<p>Interest is directed toward the provenance and contents of the dataset totaling 660 million samples used in pre-training and continued training. Questions and inquiries have been raised about what open-source datasets were utilized or how the data collection was conducted.<\/p>\n<h3>Feasibility of Local and Browser Execution<\/h3>\n<p>Given the configuration of a 2B parameter DiT and a 4B parameter text encoder (Qwen3.5-4B), discussions are taking place regarding the possibility of further model size reduction through quantization or lighter text encoders (such as replacing with T5-small or CLIP). Expectations are also expressed for an environment running directly in the browser utilizing WebGPU or similar technologies.<\/p>\n<h2>Community Reactions<\/h2>\n<h3>Technical Surprise at VAE Elimination and Expectations for High-Efficiency Models<\/h3>\n<p>Voices of surprise are rising over the fact that the approach of eliminating the separate Variational Autoencoder (VAE) previously considered standard in image generation and instead performing direct compression and generation in pixel space yields high results despite being a counter-intuitive design. It is also pointed out that because major research labs treat the scale of compute resources itself as an entry barrier and strength, incentives to focus on developing efficient models are less likely to arise. Amidst this, favorable opinions are seen praising a small-scale lab for realizing a high-efficiency model on a low budget and sharing its learning insights.<\/p>\n<h3>Challenges in Rendering Accuracy and Remarks on Sample Selection<\/h3>\n<p>On the other hand, numerous reports point out the limits of generation capabilities and low rendering accuracy. Discussions highlight that issues frequently seen in existing models persist, such as errors in hand rendering in published images, distorted limb shapes, excessive image smoothing, and unnatural glossiness. Voices are also raised expressing suspicion that the favorable generation examples published might be the result of cherry-picking good ones out of numerous trials.<\/p>\n<h3>Questions on the Origin and Content of Training Data<\/h3>\n<p>Questions are concentrated on the details of the dataset used for training. Interest gathers around specifically what open-source datasets were used or what methods were taken in collecting the data. Questions regarding data composition were observed, such as the exclusion of text-rendering data and the extent to which specific objects (like ladybugs) are included in the dataset.<\/p>\n<h3>Quantization and Browser Execution Possibilities<\/h3>\n<p>Concrete proposals toward model weight reduction and expanding execution environments have been put forward. Discussions cover behavior when quantization is applied and whether it can run directly in the browser leveraging WebGPU. Furthermore, opinions are presented suggesting that if the text encoder were replaced with a lightweight model like T5-small or CLIP and INT4 quantization applied, the overall size could be kept under 2.5GB and made runnable on a diverse range of machines.<\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/news.ycombinator.com\/item?id=50368208\">Hacker News discussion page<\/a><\/li>\n<li><a href=\"https:\/\/www.linum.ai\/field-notes\/pjit-21k\">Model explanation article by Linum AI<\/a><\/li>\n<\/ul>\n<blockquote>\n<p><strong>This article contains unverified information.<\/strong> We will append an update note once it is confirmed by a primary source.<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Explore Pyramid-JiT, a 2B text-to-image model by Linum AI trained for $21K without a VAE. Read community reactions and technical discussions.<\/p>\n","protected":false},"author":1,"featured_media":15230,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[130],"tags":[1687,1833,4676,4678,1565,3201],"class_list":["post-15231","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-community","tag-apache-2-0-en","tag-dit-en","tag-linum-ai-en","tag-pyramid-jit-en","tag-unverified","tag--en"],"lang":"en","translations":{"en":15231,"ja":15229},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15231","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=15231"}],"version-history":[{"count":1,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15231\/revisions"}],"predecessor-version":[{"id":17081,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15231\/revisions\/17081"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/15230"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=15231"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=15231"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=15231"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}