Salvato in:
Dettagli Bibliografici
Autori principali: Haji-Ali, Moayed, Menapace, Willi, Skorokhodov, Ivan, Park, Dogyun, Kag, Anil, Vasilkovsky, Michael, Tulyakov, Sergey, Ordonez, Vicente, Siarohin, Aliaksandr
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:https://arxiv.org/abs/2603.12245
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866918384518561792
author Haji-Ali, Moayed
Menapace, Willi
Skorokhodov, Ivan
Park, Dogyun
Kag, Anil
Vasilkovsky, Michael
Tulyakov, Sergey
Ordonez, Vicente
Siarohin, Aliaksandr
author_facet Haji-Ali, Moayed
Menapace, Willi
Skorokhodov, Ivan
Park, Dogyun
Kag, Anil
Vasilkovsky, Michael
Tulyakov, Sergey
Ordonez, Vicente
Siarohin, Aliaksandr
contents Diffusion transformers (DiTs) achieve high generative quality but lock FLOPs to image resolution, limiting principled latency-quality trade-offs, and allocate computation uniformly across input spatial tokens, wasting resource allocation to unimportant regions. We introduce Elastic Latent Interface Transformer (ELIT), a drop-in, DiT-compatible mechanism that decouples input image size from compute. Our approach inserts a latent interface, a learnable variable-length token sequence on which standard transformer blocks can operate. Lightweight Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions. By training with random dropping of tail latents, ELIT learns to produce importance-ordered representations with earlier latents capturing global structure while later ones contain information to refine details. At inference, the number of latents can be dynamically adjusted to match compute constraints. ELIT is deliberately minimal, adding two cross-attention layers while leaving the rectified flow objective and the DiT stack unchanged. Across datasets and architectures (DiT, U-ViT, HDiT, MM-DiT), ELIT delivers consistent gains. On ImageNet-1K 512px, ELIT delivers an average gain of $35.3\%$ and $39.6\%$ in FID and FDD scores. Project page: https://snap-research.github.io/elit/
format Preprint
id arxiv_https___arxiv_org_abs_2603_12245
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers
Haji-Ali, Moayed
Menapace, Willi
Skorokhodov, Ivan
Park, Dogyun
Kag, Anil
Vasilkovsky, Michael
Tulyakov, Sergey
Ordonez, Vicente
Siarohin, Aliaksandr
Computer Vision and Pattern Recognition
Diffusion transformers (DiTs) achieve high generative quality but lock FLOPs to image resolution, limiting principled latency-quality trade-offs, and allocate computation uniformly across input spatial tokens, wasting resource allocation to unimportant regions. We introduce Elastic Latent Interface Transformer (ELIT), a drop-in, DiT-compatible mechanism that decouples input image size from compute. Our approach inserts a latent interface, a learnable variable-length token sequence on which standard transformer blocks can operate. Lightweight Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions. By training with random dropping of tail latents, ELIT learns to produce importance-ordered representations with earlier latents capturing global structure while later ones contain information to refine details. At inference, the number of latents can be dynamically adjusted to match compute constraints. ELIT is deliberately minimal, adding two cross-attention layers while leaving the rectified flow objective and the DiT stack unchanged. Across datasets and architectures (DiT, U-ViT, HDiT, MM-DiT), ELIT delivers consistent gains. On ImageNet-1K 512px, ELIT delivers an average gain of $35.3\%$ and $39.6\%$ in FID and FDD scores. Project page: https://snap-research.github.io/elit/
title One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2603.12245