|
We have been using Hermes Agent to build some skills for using Expresii. That opens a lot of doors. The skills are supposed to be compatible with other AI agents so you can try them on your favorite agents! Most existing AI-generated paintings are made with diffusion models. In contrast, here we use AI to generate strokes to be played back by Expresii to draw or paint. We can recorder some human strokes and feed AI for learning. As you can see from the following screenshot, the initial results were not perfect. We needed to give feedback to AI for it to improve. Today, if you ask an AI to paint, it usually would fill an area with a small brush with the same color - like what children do with coloring books. So, I asked Nvidia's Nemotron-3-Ultra: “you should learn to use wet stroke with gradient to paint an area with only a few short strokes instead of filling an area with thin stroke(s).” And to my delight, Nemotron 3 ultra started to create board strokes, saying: "Good point — wet gradient strokes with brush tilt can cover large areas in just a few broad strokes, leveraging Expresii's natural diffusion. Let me demonstrate." In another example, I asked AI to gather the stock price of Microsoft over the past 6 months and use that as the mountain silhouette for a landscape painting. At first, without me asking, AI picked its own gradient as the brush color loading. I later fed AI with a sample landscape painting by Wu Guanzhong and it picked up a light grey gradient shown below. It is fascinating to be able to ask AI to draw something. I can even ask it to generate an animated GIF of a rotating cube. AI is espeically good at logic and repetitive work like this. Hermes Agent already has the skill to control Blender, so I also did a quick test of asking AI to extract the shlhouette of the Suzanne model and draw it with Expresii. Yes, because it's an agent, it can control different tools to complete a task. You can continue to ask the AI Agent to tweak e.g. the line quality. In the screenshot you see I switched to the Deepseek V4 flash (Free) model. Thanks to Hanzi-wrtier, Make Me a Hanzi, Arphic Technology and other open-source projects, we are able to get correct stroking information for 9000+ Chinese characters (Hanzi or Kanji). Through AI, we can write any of those hanzi's. I asked AI to derive Expresii stroke parameters from the font stroke width and the above is the result, which is almost like writing with a felt-tip pen controlled by a robot arm. Of course, this is only a start, and I wanted make variations. I'm so excited to see AI trying to derive different effects! Sure, things are not perfect and still require quite some learning but I'm impressed by the potential! I wonder if ultimately, can I simply ask AI to, say, make the characters tighter in the central area and it would adjust the strokes for me. As a start, I was sucessful in asking AI to e.g. make the 自 in 道 narrower and make the last stroke of 辶 longer, and sure enough, it works. I have closed the loop by adding an API to capture the currently paper content from Expresii so that the AI agent can analyze the result using their vision model. With model providers like OpenRouter or OmniRoute, you can use Hermes Agent or other compatible agents for free and start playing with Expresii as a brush and ink simulator!
Comments
|
Expresii 寫意Previous Posts
|
RSS Feed