展示HN：基于SFMTA CAD图纸训练的LoRA模型与航拍图像
嗨，我是基兰，我一直在探索生成式人工智能与土木工程在道路设计交叉的领域（这个双关语是故意的）。我进行了一个测试，使用Fal的训练器在新的Flux 2 Dev模型上训练了一个LoRA，数据集是来自公开可用的街道布局条纹CAD图纸与同一区域航拍图像的配对图像。
这个用例的目的是让城市规划者在使用现有工具时，能够即时可视化他们所提议的更改。
这只是一个小规模数据的快速实验，结果超出了我的预期，所以我想与大家分享。
观看演示并获取测试说明： [https://www.youtube.com/watch?v=zS8pGoOfe00](https://www.youtube.com/watch?v=zS8pGoOfe00)
现在就试试吧（包括免费积分）： [https://3dstreet.app/generator](https://3dstreet.app/generator)
如果你对在自己的硬件上运行感到兴奋，这里是LoRA权重：[https://v3b.fal.media/files/b/0a87f41f/glySGbKtv8lzigPWzQDjb_pytorch_lora_weights.safetensors](https://v3b.fal.media/files/b/0a87f41f/glySGbKtv8lzigPWzQDjb...)
如果你对细节感兴趣，我可以写一篇更长的博客文章。这个模型仅使用了12对图像和文本描述进行训练，但在Fal上仍花费了大约100美元。我很想进行更大规模的训练，但准备所有数据确实需要一些时间，我对投入2000美元有些犹豫。我很好奇，专家们认为如果我使用更大的样本量，质量是否会提高。
查看原文
Hey I'm Kieran and I've been playing with the intersection (pun intended) of generative AI and civil engineering for roadways. I did a test run training a LoRA on the new Flux 2 Dev model using Fal's trainer useing a custom dataset of paired images from publicly available striping CAD drawings of street layouts to aerial images of the same area.<p>The use case here is to allow urban planners to instantly visualize their proposed changes as they work with their existing tooling.<p>This was just a quick experiment with a small data size that exceeded my expectations so I wanted to share with you all.<p>Watch a demo with instructions on how to test it out: <a href="https://www.youtube.com/watch?v=zS8pGoOfe00" rel="nofollow">https://www.youtube.com/watch?v=zS8pGoOfe00</a><p>Try it now (includes free credits) <a href="https://3dstreet.app/generator" rel="nofollow">https://3dstreet.app/generator</a><p>If you're really excited about running on your own hardware here are the lora weights: <a href="https://v3b.fal.media/files/b/0a87f41f/glySGbKtv8lzigPWzQDjb_pytorch_lora_weights.safetensors" rel="nofollow">https://v3b.fal.media/files/b/0a87f41f/glySGbKtv8lzigPWzQDjb...</a><p>I can writeup a longer blog post if interested in the details. This was only trained on 12 image pairs with text descriptions but it still cost about $100 on Fal. I'd love to do a larger run, but it does take a while to prepare all of the data and I'm hesitant to drop $2k. I'd be curious for the experts out there if you think the quality will increase if I use a larger sample size.