The paper introduces Dynamics-aware Token Merging (DaTMe), a training‑free method that accelerates diffusion models by selectively merging tokens based on diffusion dynamics. DaTMe builds an importance map from the Tweedie denoised estimate, combining temporal changes and structural high‑frequency content to decide which tokens to preserve and which to merge. Experiments on PixArt-a and FLUX show that DaTMe maintains visual fidelity across the entire image while reducing token computation.
arXiv:2608.23864v1 Announce Type: new Abstract: Visual tokenizers increasingly inject semantic supervision into latent spaces to make downstream diffusion easier. Yet how these semantics should be or...
arXiv:2608.24293v1 Announce Type: new Abstract: Latent diffusion models have emerged as a dominant framework for high-fidelity image and video synthesis, operating in compact latent spaces with varia...
arXiv:2608. 03135v1 Announce Type: cross Abstract: Text-to-image diffusion models can generate individual concepts well, but they often omit or merge concepts incorrectly with multiple concepts.
arXiv:2505. 18227v4 Announce Type: replace-cross Abstract: In Transformer architectures, tokens\textemdash discrete units derived from raw data\textemdash are formed by segmenting inputs into fixed-length chunks.
arXiv:2606. 20076v1 Announce Type: cross Abstract: Latent Diffusion Models (LDMs) have become dominant in visual synthesis, but their quality-compute trade-off is largely constrained by the tokenizer's fixed compression ratio.
arXiv:2608. 01821v1 Announce Type: cross Abstract: Diffusion vision-language models (dVLMs) iteratively denoise masked responses while conditioning each denoising step on visual evidence, making visual conditioning a substantial recurring inference cost.
arXiv:2606. 13289v1 Announce Type: cross Abstract: Holistic visual tokenizers are fundamental to unified multimodal models (UMMs) as they map diverse visual inputs into a unified representation space.
The paper proposes InterpTok, a method that fine‑tunes the MAETok image tokenizer to better handle codes that lie between training samples. By training the decoder on mixtures of image codes and using a discriminator plus a mixing‑ratio loss, InterpTok preserves reconstruction quality while dramatically improving interpolation and generation metrics. The approach yields sharper, more natural image compositions and significantly lowers FID scores for diffusion models trained on the new tokenizer.
The paper argues that token importance alone is insufficient to determine safe removal of visual tokens in multimodal large language models, because removability depends on representation depth and the surrounding deletion set. Through controlled experiments, the authors show that the same tokens can have different effects when removed at different depths or contexts. They introduce CoRePrune, a training‑free two‑stage pruning framework that refreshes deletion effects as visual representations evolve and refines candidate tokens based on the current deletion set, achieving high performance retention across multiple backbones and reducing prefill time significantly.
BudgetPix introduces a compute‑adaptive tokenization framework for pixel‑space image diffusion, dynamically allocating computational resources based on visual complexity and spatial layout. It comprises an adaptive encoder using an entropy‑guided quadtree and multi‑scale patch embedder, a scale‑aware decoder, and a flexible training schedule that allows pixel‑space denoisers to operate across variable token counts. The method integrates with existing diffusion architectures, achieving comparable fidelity to larger models while using only a fraction of the compute budget, and demonstrates improved quality‑efficiency tradeoffs in both text‑to‑image and class‑conditional generation.
Visual Information-Guided Parallel Decoding for Diffusion Multimodal Large Language Models introduces the VIG‑Sampler, a method that prioritizes tokens for decoding based on their attention to image tokens and penalizes redundancy in image‑attention distributions. The approach aims to improve the quality of multimodal generation by selecting more informative tokens during diffusion decoding. Experiments on seven captioning and VQA benchmarks with three open‑source dMLLMs show that VIG‑Sampler outperforms the Info‑Gain Sampler by an average of 19.3 CIDEr points and achieves better COCO Caption results using only half as many decoding steps.
arXiv:2610.00686v1 Announce Type: new Abstract: Recent video-based world models pair the scalability of autoregressive (AR) prediction with the visual quality of diffusion models. The choice of scene...