Image generators can add detail while largely preserving image structure

Across three image generators, the authors demonstrate a training-free detail control that adjusts internal features rather than relying on phrases such as “highly detailed”.

Big Tech

Qingtao Yu · Zhen Wang · Jiyang Zheng · Minghao Sun · Hongdong Li · Dylan Campbell

The Australian National University · Microsoft · The University of Queensland

Research Digest·3 November 2026·2 min read

Illustration: Image generators can add detail while largely preserving image structure

In one line

Flattening a token's energy distribution by scaling its channel magnitudes gives continuous control over generated image detail.

Before you read this

Three ideas this paper builds on

diffusion transformerA neural network architecture combining transformer layers with diffusion processes for high-quality image and video generation.65 papers →activation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.57 papers →mechanistic interpretabilityThe study of how neural networks compute through internal circuits, features, and causal interventions, rather than just input-output behavior.104 papers →

Read it as

Text-to-image models often struggle to produce pictures with just the right amount of detail. Adding phrases like "highly detailed" to the prompt can help, but the effect is unpredictable and hard to fine-tune. This paper finds a better way.

The authors discovered that when a prompt asks for more detail, the internal representations of the model become flatter: the energy that was concentrated in a few channels spreads out more evenly across many. They call this flattening effect.

Using this insight, they created a small piece of code that artificially flattens those representations by a user-chosen amount. This gives continuous, predictable control over image detail without retraining the model. It preserves image quality and semantic meaning, and adds almost no extra computation time.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

token flattening diffusion transformer flow matching activation steering classifier-free guidance mechanistic interpretability