Omni‑Diffusion‑Distill is a two‑stage distillation framework that compresses unified multimodal diffusion large language models (dLLMs) into a single efficient student while preserving both image generation and multimodal understanding. In the first stage the student learns to skip decoding steps by replaying cached teacher trajectories, and in the second stage it is refined on intermediate states along its own rollouts. The method introduces a pairwise collision penalty to reduce repetition in parallel text decoding and entropy‑matched guidance to prevent entropy collapse in image generation, achieving state‑of‑the‑art trade‑offs: reducing image generation from 128 to 8 steps and multimodal understanding from 512 to 64 steps, yielding 18.2× and 21.2× wall‑clock speedups and strong performance on GenEval, DPG‑Bench, MM‑Vet, and COCO captioning.
arXiv:2604. 16514v5 Announce Type: replace-cross Abstract: Autoregressive vision-language models (VLMs) deliver strong multimodal capability, but their token-by-token decoding imposes a fundamental inference bottleneck.
arXiv:2610.19305v1 Announce Type: new Abstract: One-step diffusion distillation achieves strong performance on small datasets but remains insufficient for high-resolution or complex text-to-image set...
The paper introduces Phase-wise Velocity Distillation (PVD), a method that splits the diffusion generation timeline into coarse and fine phases, each handled by a half-sized expert. By decoupling structural composition from detail refinement, PVD achieves higher quality image generation within a single full-backbone compute budget, outperforming monolithic student models. Experiments on ImageNet and text-to-image tasks show competitive FID scores and significant reductions in parameters and VRAM usage.
Unified multimodal models (UMMs) interleave generated ''visual thoughts'' (VTs) with text reasoning to improve spatial tasks. This incurs roughly an order-of-magnitude inference cost from multi-step diffusion.
The paper introduces VDOT++, a unified distillation framework for few-step video generation across text-to-video, image-to-video, and condition-based tasks. It extends the previous VDOT method by incorporating an asymmetric unbalanced optimal transport distillation that allows student tokens to carry less mass, and an α1 ground cost that uses a weighted median to preserve modes. The approach combines distribution matching with adversarial refinement and cross-scale distillation, achieving competitive performance on multiple video benchmarks with only four sampling steps.
arXiv:2601. 09881v2 Announce Type: replace-cross Abstract: Large video diffusion and flow models have achieved remarkable success in high-quality video generation, but their use in real-time interactive applications remains limited due to their inefficient multi-step sampling process.
Intern Lumina U2 is a diffusion large language model that uses multi-codebook visual tokens in a fully discrete space, allowing each visual position to be represented by a group of complementary codes. The model concatenates embeddings at each position and employs a lightweight autoregressive head to predict codes sequentially while the diffusion backbone processes all positions in parallel. Built on the unified 4D representation of AToken, it supports image, video, and 3D inputs, achieves leading performance in visual understanding, image generation, and editing, and includes a cross‑platform training, inference, and evaluation system with a stability‑aware sampling strategy that speeds up image generation by 4.45×.
Diffusion Language Models (DLMs) have recently achieved substantial progress in natural language generation tasks. Recent research demonstrates that adaptive token generation ordering can significantly improve performance in mathematical reasoning and code synthesis applications.
UltraDiff is a new framework for high‑resolution image generation with Latent Diffusion Models that tackles the trade‑off between computational cost and fidelity caused by latent compression. It jointly learns latent diffusion and then distills a marginal score using a dual‑branch DiT called MixTeacher, which models both f8 and f32 distributions and resolves compressed‑space ambiguity with high‑fidelity priors. Vector‑field distillation in the native f32 space forces the student model to learn the teacher’s conditional expectation, filtering out compression artifacts and achieving f8‑level quality while retaining f32 efficiency—generating a 2K image in 6.2 seconds on a single A100 GPU, over 40× faster than the f8 baseline.
arXiv:2607. 14557v1 Announce Type: new Abstract: Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints.
arXiv:2609.40362v1 Announce Type: new Abstract: We present Multimodal Flow, a fully continuous generative model of language and vision. Most unified multimodal models either model both language and q...
arXiv:2607. 08056v1 Announce Type: cross Abstract: Diffusion Language Models (DLMs) have recently achieved substantial progress in natural language generation tasks.