Linum团队发布了JiT-DDT架构，一种改进的文本到图像模型训练方法1。相比Linum v2基线，该方法将训练速度提升3.6倍（以GPU小时数计），同时生成的图像像素数增加为4倍1。
JiT-DDT通过采用x-prediction和v-loss流匹配等技术创新来克服像素空间模型在生成精细细节方面的难题1。其架构采用编码器-解码器设计，其中编码器负责预测8倍下采样图像（从512×512降采样至64×64），解码器则处理高频细节以生成全分辨率图像1。该团队还引入了PixelREPA辅助损失、SwiGLU激活函数、Muon优化器和Qwen3.5-4B文本嵌入等多项改进措施1。
代码和模型权重已在Apache 2.0许可证下开源发布，作为研究工件形式推出1。该团队由两位兄弟组成，致力于从零开始训练文本到视频模型，目标是使动画制作更加易于获取1。
A research team has introduced JiT-DDT, an improved architecture for training text-to-image models that achieves substantially faster training times while generating higher-resolution outputs 1. The new approach trains 3.6 times faster in terms of GPU hours compared to Linum v2 baseline, while producing images with 4 times more pixels 1.
The architecture employs several technical innovations to overcome limitations of pixel-space models in generating fine details 1. Rather than using traditional v-prediction with v-loss flow matching, JiT-DDT implements x-prediction and v-loss flow matching 1. The design separates processing into an encoder stage that predicts an 8-fold downsampled image (reducing 512×512 resolution to 64×64) and a decoder stage that handles high-frequency details to generate the full-resolution final image 1. Additional improvements include the PixelREPA auxiliary loss, SwiGLU activation functions, the Muon optimizer, and Qwen3.5-4B text embeddings 1.
The code and model weights have been released under the Apache 2.0 license as research artifacts rather than as a complete production model 1. The research is being conducted by a team of two brothers working to train text-to-video models from scratch, with the goal of making animation production more accessible 1.
评论
还没有评论，欢迎留下第一条。