arXiv:2511.10629v2 Announce Type: replace 
Abstract: Diffusion models struggle to scale beyond their training resolutions, as direct high-resolution sampling is slow and costly, while post-hoc image super-resolution (ISR) introduces artifacts and additional latency by operating after decoding. We present the Latent Upscaler Adapter (LUA), a lightweight module that performs super-resolution directly on the generator's latent code before the final VAE decoding step. LUA integrates as a drop-in component, requiring no modifications to the base model or additional diffusion stages, and enables high-resolution synthesis through a single feed-forward pass in latent space. A shared Swin-style backbone with scale-specific pixel-shuffle heads supports 2x and 4x factors and remains compatible with image-space SR baselines, achieving comparable perceptual quality with nearly 3x lower decoding and upscaling time (adding only +0.42 s for 1024 px generation from 512 px, compared to 1.87 s for pixel-space SR using the same SwinIR architecture). Furthermore, LUA shows strong generalization across the latent spaces of different VAEs, making it easy to deploy without retraining from scratch for each new decoder. Extensive experiments demonstrate that LUA closely matches the fidelity of native high-resolution generation while offering a practical and efficient path to scalable, high-fidelity image synthesis in modern diffusion pipelines.

تم تقديم محول التكبير الكامن (LUA) كعنصر خفيف يعزز نماذج الانتشار من خلال إجراء تكبير فائق مباشرة على الشيفرة الكامنة للمولد قبل خطوة فك التشفير النهائية VAE. تتيح هذه الابتكار توليد صور عالية الدقة من خلال تمريرة واحدة فقط للأمام في الفضاء الكامن، مما يقلل بشكل كبير من وقت فك التشفير والتكبير مقارنة بالطرق التقليدية.

Se ha presentado el Adaptador de Escalado Latente (LUA) como un módulo ligero que mejora los modelos de difusión al realizar super-resolución directamente en el código latente del generador antes del paso final de decodificación VAE. Esta innovación permite la síntesis de imágenes de alta resolución a través de un único paso de avance en el espacio latente, reduciendo significativamente el tiempo de decodificación y escalado en comparación con los métodos tradicionales.

L'adaptateur d'upscaling latent (LUA) a été présenté comme un module léger qui améliore les modèles de diffusion en effectuant une super-résolution directement sur le code latent du générateur avant l'étape finale de décodage VAE. Cette innovation permet une synthèse d'images haute résolution grâce à un seul passage en avant dans l'espace latent, réduisant considérablement le temps de décodage et d'upscaling par rapport aux méthodes traditionnelles.

The Latent Upscaler Adapter (LUA) has been introduced as a lightweight module that enhances diffusion models by performing super-resolution directly on the generator's latent code before the final VAE decoding step. This innovation allows for high-resolution image synthesis through a single feed-forward pass in latent space, significantly reducing decoding and upscaling time compared to traditional methods.

One Small Step in Latent, One Giant Leap for Pixels: Fast Latent Upscale Adapter for Your Diffusion Models

Was this article worth reading? Share it

LucidQuery AI

SwapAnything.io

ImgUpscaler AI

CreativePixel

The Visualizer

Sprello

Ready to build your own newsroom?