Efficient Resolution-Agnostic
Generation & Editing
News
ECCV 2026 Jun 2026 UltraImageGen accepted to ECCV 2026 — Efficient Ultra-High-Resolution Image Generation with Hierarchical Local Attention. CVPR 2026 Mar 2026 HierEdit accepted to CVPR 2026 — Region-Aware Hierarchical Diffusion for Efficient High-Resolution Editing.

Efficient. Scalable. Resolution-Agnostic.

Pushing the Limits of Ultra-High-Resolution Generation and Editing

01 ECCV 2026

UltraImageGen

Efficient Ultra-High-Resolution Image Generation with Hierarchical Local Attention

UltraImageGen generates ultra-high-resolution images up to 8K × 8K with a hierarchical local attention mechanism, preserving fine-grained details efficiently.

Code Paper Abstract · arXiv 2510.16325 Ultra-high-resolution text-to-image generation is increasingly vital for applications requiring fine-grained textures and global structural fidelity, yet state-of-the-art models such as FLUX and SD3 remain confined to sub-2MP resolutions due to the quadratic complexity of attention and the scarcity of high-resolution training data. We present UltraImageGen, a framework that introduces hierarchical local attention with low-resolution global guidance, enabling efficient, scalable, semantically coherent synthesis at ultra-high resolutions. High-res latents are split into hardware-aligned fixed-size local windows to reduce attention from quadratic to near-linear, while a low-res latent with scaled positional embeddings injects global semantics as an anchor. A lightweight LoRA bridges global and local pathways, ensuring consistency across structure and detail. UltraImageGen reliably scales pretrained models beyond 8K with >10× speed-up and significantly lower memory, achieving superior quality while maintaining efficiency.
UltraImageGen results — ultra-high-resolution generated scenes
02 CVPR 2026

HierEdit

Region-Aware Hierarchical Diffusion for Efficient High-Resolution Editing

HierEdit enables high-fidelity local editing and inpainting at ultra-high resolutions with region-aware hierarchical diffusion, supporting both instruction editing and text-/image-guided inpainting.

Code Paper Abstract · arXiv 2605.17294 High-resolution image editing is essential for professional and creative applications, yet existing multimodal diffusion-based editors remain computationally inefficient and constrained to low resolutions. Current approaches redundantly process the entire image canvas or rely on large-scale high-resolution datasets, resulting in substantial training and inference costs. We introduce HierEdit, a region-aware hierarchical diffusion framework for efficient, scalable high-resolution editing. Our method first edits a low-resolution proxy using an off-the-shelf editor to generate a reference and localize modified regions. A hierarchical local-window MMDiT then refines only edited regions within the original high-res image while reusing the unaltered regions as conditioning. The low-res proxy further provides structural guidance and intermediate denoising supervision, ensuring consistent global semantics without full-resolution attention. HierEdit enables fast, high-fidelity editing up to 4K without specialized high-res training data, while significantly accelerating inference.
HierEdit results — region-aware high-resolution editing
03 Under Review

AtlasVid

Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling

AtlasVid generates ultra-high-resolution long videos up to 8K and hundreds of frames with decoupled global-local modeling, achieving strong efficiency and visual quality.

Coming Soon Paper Abstract · arXiv 2605.16649 Recent diffusion-based video generators have achieved remarkable visual fidelity and prompt controllability, yet scaling them to ultra-high-resolution (UHR) long videos remains prohibitively expensive. We present AtlasVid, which decouples global and local modeling: a low-resolution, low-FPS global semantic proxy is generated via temporally scaled RoPE, while a high-resolution detail branch performs joint denoising within local spatiotemporal windows. Trained in a resolution-agnostic fashion at 720P, AtlasVid generalizes to 4K and beyond for videos exceeding 10 seconds, achieving a 60.9× speed-up over native 4K generators while maintaining superior visual quality and requiring significantly reduced training overhead.
AtlasVid results — long ultra-high-resolution video generation