New AI System Teaches Itself to Make Images, No Retraining Needed

Most AI image tools work one job at a time. You type a request, it makes a picture, and tomorrow it starts from scratch with no memory of what worked. OmniHarness, a new research system from Beihang University, the Chinese University of Hong Kong and the National University of Singapore, changes that by getting the AI to keep a written playbook. Every time a task succeeds, the system saves the general recipe — the steps and the conditions where they apply — like a chef writing down a dish so the kitchen can repeat it without guessing.

The AI also checks its own work partway through, not just at the end, so it can notice a mistake and change course mid-task. And it rehearses on its own, inventing practice jobs near the edge of what it can do, the way a musician runs scales before a concert. The clever part: the AI's underlying brain stays frozen. All the improvement lives in the recipe book, which means no expensive retraining and recipes that can be plugged into other AI systems as-is.

The results are striking on paper. On ComfyBench's creative tasks, OmniHarness completed 95% of requests, beating the strongest previous system by 27.5 percentage points. It was tested across six benchmarks, three different AI models and three agent frameworks — agents being AI that can take actions on your behalf. The catch: this is an academic paper with no public product, the tests measure whether the AI finished the requested workflow, not whether the picture looks good, and narrow tool-use benchmarks rarely capture real creative judgment.

If this reaches real products, expect image and video editing tools that remember your style and past edits, cheaper AI services because nobody has to retrain a model, and far less time wasted re-typing the same instructions. That's years away, but the direction — AI that learns from its own experience without a costly rebuild — is the one every major lab is chasing.