正在加载视频...

视频加载失败

This is native Qwen Image 2.1, no LoRA - move and resize objects seamlessly in an image🔥 I was planning to finetune on this task to showcase the new 🧨 diffusers LoRA support for Qwen 2.1 (go try it!), but while testing the base model I realized it's actually...

53,464 次观看 • 1 个月前 •via X (Twitter)

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

it got worse? Nano Banana 2.1 is here so had to put it up against NB2 for the good old continuity stress test. the strange thing is 2.0 kept my likeness a lot more all the way through. I'm still very different from the original, but you can kinda recognise me? On 2.1 it just regenerates a new man multiple times. I guess they've made changes to the Synth ID watermarking they do on the images and the noise causes the model to go blind? it's funny how both models seem to become conscious of the fact they've made a complete mess with the red distortion, and then revert back to a clean generation with no noise at all. but when 2.1 does it, it completely forgets what I look like because the distortion is heavier on 2.1. Looking at the positives though 2.1 is it's much better at knowing what needs to be removed even with a super simple prompt. In 2.0, after the flashlight, it kept the light beams in the generations after that. then holding the teacup, it brought in another hand to hold the bottom of the cup and that hand didn't disappear in the generations afterwards. Whereas on 2.1, my position and posture never changed. From one generation to the next, the entire object disappeared and the new one was placed exactly where the previous one was. No drift in that sense. so 2.1 is the better editor. 2.0 is better at remembering what I look like. I think watermarking is a great, but it needs to go unnoticed for at least 10 iterations on the same image. It's too quick to notice in NB2.1. Still 1 or 2 iterations better than GPT Image 2.5 though.

Alec Wilcock

32,004 次观看 • 1 个月前