SEGA helps diffusion models output 4K images without retraining

Anyone who generates AI images has hit this wall: an illustration looks crisp at default dimensions, but bump the canvas to 4K for a poster or wallpaper, and the model completely loses its mind—sprouting extra limbs or turning textures into repetitive mush.

Diffusion transformers (DiTs) have taken over modern image synthesis, but they break down hard once pushed past their native training resolutions. Their positional awareness simply cannot handle the extra real estate.

To tackle this bottleneck, researchers introduced SEGA (Spectral-Energy Guided Attention) in a paper accepted to NeurIPS 2026. The pitch is straightforward: it is a completely training-free method that lets DiT models output clean 4096×4096 visuals without retraining or expensive fine-tuning.

Ditching the one-size-fits-all scaling

Prior training-free fixes, including methods like YaRN, DyPE, and UltraImage, tried tweaking Rotary Position Embeddings (RoPE) extrapolation alongside static attention scaling.

Their main flaw was treating all frequencies the same.

Images are naturally split between low spatial frequencies that dictate composition and high frequencies that handle textures like fur, pores, and ripples. When an inference method applies uniform, content-agnostic scaling across every RoPE dimension, it forces an ugly compromise: preserve the rough layout and blur the details, or push the details and watch the global anatomy fall apart.

SEGA swaps fixed scaling for an adaptive mechanism that watches what is actually happening inside the image.

Reading frequencies on the fly

At every denoising step, the DiT model processes a noisy latent representation. SEGA reads this representation directly without adding external neural modules or learned weights.

It reshapes the latent into its spatial layout and runs a fast Fourier transform (FFT) to measure how energy is distributed across different frequency bands.

Because individual RoPE dimensions correspond to specific spatial frequencies, SEGA maps that spectral energy directly back onto the position embeddings. Frequency bands with lower energy get stronger scaling to reconstruct missing structural bounds, while high-energy bands receive lighter scaling to prevent artifacts from over-amplification.

This scale dynamically shifts across dimensions, spatial axes, denoising steps, and individual prompts.

Tested on models like Flux and Qwen at 4096×4096, SEGA generated intricate prompts—such as an underwater polar bear surrounded by bioluminescent plankton or the hyper-detailed scales of a Nile crocodile eye—without the typical structural melt seen in baseline methods.

Why it matters

For digital illustrators, visual designers, and marketers, scaling AI artwork to direct-to-print resolutions has always been a painful bottleneck. Most workflows still rely on generating smaller frames and passing them through heavy upscaling pipelines, which often distorts subtle aesthetics.

SEGA demonstrates that you do not need massive compute budgets or post-training runs just to unlock 4K output. Because it operates purely at inference time, open-source developers can drop the mathematical logic directly into existing sampling loops without changing weights.

We think bridging classical signal processing like FFT with modern transformer mechanics is a remarkably clever move. Instead of treating the DiT purely as a black box, using spectral energy to steer RoPE directly fixes positional drift where it actually starts.

There are tradeoffs, of course. Calculating frequency maps at every sampling step introduces slight inference overhead, and the method cannot invent stylistic depth that the base model never possessed. Still, for teams looking to deliver high-resolution media straight from current models, this is a very practical upgrade to the generative toolkit.