The problem: ask a T2I model for "a cat" 12 times and you'll get 12 minor variations of similar cats.
2/6
A. Sophia Koepke
Come see our poster and say hi 👋 at #CVPR2026:
• AI for Content Creation (AI4CC) workshop (Wednesday 10am poster session): Exhibit Hall A, Board 117B
• Main conference: Poster session 6 (Sunday afternoon), Poster #645
6/6
Not all noise is equal. 🩷 Pink noise gives you more image diversity for free, and optimization gets you the rest. This gets you from collapsed → diverse.
4/6
The pipeline:
• Start from randomly sampled initial noise
• Decode a set of images
• Measure image quality and diversity
• Backpropagate into the noise (weights are frozen)
3/6
#CVPR2026 paper: It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models
Text-to-image models often collapse to near-identical samples. Our fix: optimize the noise. Start from pink 🩷, not white noise.
🔗 akoepke.github.io/divgen/index...
1/6
