Urgent.News

What's breaking now, across thousands of outlets.

AI

Your AI Character Looks Right—Except Every Asymmetric Detail Flipped

This article comes from the author's work maintaining Tsuduri, a recurring AI-illustrated character with asymmetric design features. The author defined and verified the workflow in Japanese; AI recomposed it for DEV readers and redrew the workflow diagram. #ABotWroteThis Tsuduri's hair ornament belongs on her left side. When she faces the viewer, that ornament appears on the right side of the…

This article discusses an issue with an AI-illustrated character named Tsuduri. Tsuduri is designed with asymmetric details, meaning some features will appear on different sides depending on the character's pose. For example, her hair ornament is on her left side when facing the viewer, but on the right side when facing away. However, an image model may incorrectly place the ornament on the wrong side, making the character look like a different character altogether.

The author addresses this problem by implementing a workflow that keeps a short invariant in the character's own coordinate system. This invariant states that the hair ornament is attached to the character's left side. The workflow then starts with the single reference view closest to the requested composition, classifies the actual output view, and validates at least two asymmetric traits.

The article emphasizes the importance of separating viewer-relative instructions from object-relative ones. While the image may look correct, the wrong character state is still present. The author suggests that the single reference view should be screen-relative, and the invariant should be kept regardless of the camera position or character pose. This approach ensures that the character's left side always appears on the left side of the image, regardless of the pose or camera angle.

The author concludes by stating that following this workflow will lead to more accurate and consistent illustrations, even if an image looks beautiful. Image quality and spatial correctness are separate checks, and the solution to this problem lies in using a closed loop that inspects the output and corrects it as needed.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

The Dedicated OCR Engine Lost to the General-Purpose Model — 300 Slower

Originally published on hexisteme notes . I had a 27B vision model running locally (IQ4_XS quantized, 15GB resident) and needed to decide whether it was worth using for OCR.

  • Dedicated OCR engine processed image in 82.8 seconds
  • General-purpose model completed task in 0.27 seconds
  • General-purpose model maintained document structure better

More from Saturday 5 September →