Training-free method that rescales attention across RoPE components to generate ultra-high-resolution images with diffusion transformers.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
SEGA (Spectral-Energy Guided Attention) is a research method for resolution extrapolation in diffusion transformers (DiTs). It lets a pretrained text-to-image DiT generate images well beyond its training resolution, up to about 36 megapixels, without retraining or adding learned parameters. The paper is from Rajabi, Shaban, Roohi, Lindell and Taati and is listed for NeurIPS 2026.
At each denoising step, SEGA takes an FFT of the current latent to estimate spatial-frequency energy. It maps that energy onto the RoPE dimensions, which encode specific spatial frequencies, and rescales attention per dimension. Low-energy bands get stronger scaling to recover under-resolved structure, and high-energy bands get weaker scaling to avoid over-amplification. Earlier training-free approaches apply a uniform, content-agnostic scaling, which trades global structure against fine detail.
The project page compares SEGA with YaRN, DyPE and UltraImage at 4096×4096 on Flux and Qwen-Image. A YouTube creator reported that code was released for Flux 1 and Qwen Image, but no repository link appeared in the provided material. It is a technique for people who run these diffusion models, not a standalone product or hosted service.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Method for generating very high-resolution images (4K to 6144 px) with sharp detail, working with Flux and Qwen Image bases. More consistent than other upscalers such as DyPE; code released for Flux 1 and Qwen Image.
Closest neighbours by meaning, not just by category.
Model · NVIDIA
NVIDIA research system where a code game engine drives a two-step streaming MiniMax-H3 video model for real-time rendering.
Model
Streaming real-world video super-resolution using sparse large-model keyframes and a lightweight propagation network.
Model · ByteDance
ByteDance's open-weights one-step diffusion transformer for restoring and upscaling video and images.
Research · ByteDance
ByteDance research method for post-training image diffusion models using on-policy self-distillation from reward signals.
Research
Multi-view video diffusion model that turns monocular or sparse videos of dynamic humans into dense multi-view videos for 4D Gaussian splats.
Model
Research model for real-time, audio-driven talking-head video generation from a reference image, using one-step inference.