Understand AI community's frustration with first impressions of #stablediffusion2, I was too. But after experimented new prompt engineering with various styles by SD2+Diffusers runned in DFserver, I discovered advantages and uniqueness of SD2 @EMostaque@RiversHaveWings 1/n
The thread all generated by pure prompt, no init image, no editing. View full prompt in alt of each images.
Cons of SD2: in addition to higher resolutions, significant improvements in image quality and detail richness, especially in handling light sources, shadow, reflection 2/n
Fineness of texture details of surface from various materials is awesome. I tried sand, snow field, ocean wave foam, liquid metals. 3/n
For prompt enginering of SD2, either too short or too long prompt doesn't work well. I found SD2 can offer higher controllability with more precision than 1.5 So highly targeted prompt design is possible, not stacking of keywords or artist refer as blindfold alchemy anymore. 4/n
Experiment with Color Palette, accurate and artistic outcomes.
blue & red | Yellow & blue | black & white | teal & burnt sienna
The anatomical coherence of human face is good, I picked below 4 from less than 20 generations.
Experiment with painting medium, oil vs watercolor.
The expressiveness of brush strokes and edge rendering, also the simulation of canvas surface for different medium is stunning.
the environmental reflections on the surface of Glass and copperware 😊
Prompt & seed in ALT
6/n
Another experiment with fine art painting medium, oil vs watercolor for landscape, in style of Andreas Rocha (although he is a digital painting master)
Prompt & seed in ALT
7/n
Heard that celebrities' face feature was removed from 2.0 dataset, but I think can be easily made up by customed finetune. As foundational open model, It's right to give more consideration to ethical controversies. I tried printmaking portraits, surely can guess who they are. 8/n
I organized this thread into a more complete article, which contains more tips and experiment results for sharing, plus progressive optimization process of 2 sets of prompts. 9/n
强烈推荐一个 AI 主题英文播客,Eye on AI。主持人Craig是纽时多年资深记者,现美国国家人工智能安全委员会(NSCAI)特别雇员。
最新一期嘉宾 ILYA SUTSKEVER (openAI 首席科学家+ChatGPT之父),访谈质量非常之高。
这个AI播客经营了5年,早早领先于这波热潮,还有一份双周 research newsletter。 1/3
Eye on AI 访谈过的嘉宾除了 深度学习三巨头里的Geoff Hiton和杨立昆。 «计算机体系架构»作者、图灵奖得主和 TPU开创者 David Patterson, 还有大量一线AI研究者和相关领域学者,涉及了AI会影响的当下及未来的方方面面。
所有播客都能下载文字版本,原文困难的同学用GPT翻译下,学英语的好材料。 2/3
另外我想推荐一个中文公众号OneFlow,链接是他们经授权后翻译的 Eye on AI 跟 ILYA SUTSKEVER 的对谈。虽然 OneFlow是中国一家做ML框架和实训社区的纯技术公司,我也是意外发现他们的媒体团队应该是中文自媒体里采编译水平最高的,无论内容源还是翻译质量,技术方面也很过硬 mp.weixin.qq.com/s/rZBEDlxFVsVX…