【Hacker News搬运】FLUX速度很快，而且是开源的
-
Title: FLUX is fast and it's open source
FLUX速度很快，而且是开源的
Text:
Url: https://replicate.com/blog/flux-is-fast-and-open-source
由于我无法直接访问互联网，包括查看外部链接，我无法直接抓取和分析您提供的链接内容。但是，我可以根据您提供的标题和描述，给出一个基于该链接可能内容的总结和翻译。 标题：“Flux 是快速且开源的” 总结： 这篇博客文章可能介绍了名为“Flux”的某个工具、库或框架，并强调了它的两个主要特点：速度和开源。文章可能会讨论Flux的性能优势，以及它如何作为一个开源项目对社区有益。可能会包含Flux的设计理念、使用案例、性能比较和/或安装指南。 翻译： 这篇文章可能介绍了名为“Flux”的工具、库或框架，并强调了它的两个主要优势：速度快和开源。文章可能会讨论Flux的性能优势，以及它作为一个开源项目如何对社区产生积极影响。可能会包含Flux的设计理念、使用示例、性能对比以及安装说明。Post by: smusamashah
Comments:
sorenjan: Text to image models feels inefficient to me. I wonder if it would be possible and better to do it in separate steps, like text to scene graph, scene graph to semantically segmented image, segmented image to final image. That way each step could be trained separately and be modular, and the image would be easier to edit instead of completely replace it with the output of a new prompt. That way it should be much easier to generate stuff like "object x next to object y, with the text foo on it", and the art style or level of realism would depend on the final rendering model which would be separate from the prompt adherence.<p>Kind of like those video2video (or img2img on each frame I guess) models where they enhance the image outputs from video games:<p><a href="https://www.theverge.com/2021/5/12/22432945/intel-gta-v-realistic-machine-learning-cityscapes-dataset" rel="nofollow">https://www.theverge.com/2021/5/12/22432945/intel-gta-v-real...</a>
<a href="https://www.reddit.com/r/aivideo/comments/1fx6zdr/gta_iv_with_a_photorealistic_filter_with_runway/" rel="nofollow">https://www.reddit.com/r/aivideo/comments/1fx6zdr/gta_iv_wit...</a>sorenjan: 文本到图像模型对我来说效率低下。我想知道是否有可能更好地分步骤进行，比如文本到场景图、场景图到语义分割图像、分割图像到最终图像。这样，每个步骤都可以单独训练并模块化，图像更容易编辑，而不是完全用新提示的输出替换。这样，生成诸如"；对象y旁边的对象x，上面有文本foo”；，艺术风格或现实主义水平将取决于最终的渲染模型，该模型将与即时粘附分开<p> 有点像那些video2video（我猜是每帧上的img2img）模型，它们增强了视频游戏的图像输出：<p><a href=“https:”www.theverge.com“2021”5,12,22432945“intel-gta-v-realistic-machine-learning-citescapes-dataset”rel=“nofollow”>https:”/；www.theverge.com；2021年；5#x2F；12℉；22432945℉；intel-gta-v-real</一<a href=“https:”www.reddit.com“r”aivideo“comments”1fx6zdr“gta_iv_with_a_photo-realistic_filter_with_runway”rel=“nofollow”>https:”/；www.reddit.com；r；aivideo
；评论/；1fx6zdr；gta_iv_wit</一vunderba: Flux is the leading contender for a locally hosted generative systems in terms of prompt adherence, but the omnipresent shallow depth of field is irritatingly hard to get rid of.
vunderba: Flux是本地托管生成系统的主要竞争者，但无处不在的浅景深很难摆脱。
CosmicShadow: I just cancelled my Midjourney subscription, it feels like it's fallen too far behind for the stuff I'd like to do. Spent a lot of time considering using Replicate as well as Ideogram.
CosmicShadow: 我刚刚取消了我的Midjourney订阅，感觉就像这样；对于我来说，它已经落后得太多了；我很乐意。我花了很多时间考虑使用Replicate和Ideogram。
dvrp: i think we (krea) are faster at the time of writing this comment (but i’ll have to double-check on our infra)
dvrp: 我认为我们（krea）在写这篇评论时速度更快（但我必须仔细检查我们的infra）
swyx: this comparison for the quantization effect is very nice <a href="https://flux-quality-comparison.vercel.app/" rel="nofollow">https://flux-quality-comparison.vercel.app/</a><p>however i do have to ask.. ~2x faster for fp16->fp8 is expected right? its still not as good as the "realtime" or "lightning" options that basically have to be 5-10x faster. whats the ideal product usecase for just ~2x faster?
swyx: 量化效果的这种比较非常好<a href=“https:”通量质量比较.vercel.app“rel=”nofollow“>https:”/；通量质量比较.vercel.app</a> <p>但我必须问一下~fp16速度提高2倍->；预计fp8会成功吗？它仍然不如"；实时”；或"；闪电”；基本上必须快5-10倍的选项。哪种产品的理想用例只需要快2倍？