Looking at the “woman, messy hair, high quality” image, the hair farther from her head looks wrong in much the way that iPhone portrait mode messes up hair. I wonder if this is an example of an AI training on partially AI-generated data and reproducing its artifacts.
Paper: <a href="https://arxiv.org/abs/2402.17113" rel="nofollow">https://arxiv.org/abs/2402.17113</a><p>The author Lvmin Zhang is the same person behind ControlNet.
The partial alpha blending support for translucent materials is really cool (glass, plastic, etc).<p>I'd be curious to see how well this plays with inpainting. Apparently img2img is also on the authors todo list.
reactions<p>1 - the way the dog at the end gets a reflection off the floor is pretty nice.<p>2 - i wonder how this compares in terms of latency/complexity with a comfyui pipeline that just does a typical edge detection/masking layer to achieve the transparency effect. however i dont think that method would work with the glass example as shown