Diffusion Models

GroundAnything treats visual grounding as joint visual-evidence extraction rather than left-to-right generation. Its 4B-parameter model uses bidirectional diffusion and explores up to 4.51× faster inference.

SoloQ is a calibration-free weight-and-activation quantization framework designed for diffusion language models. It combines structured rotations, rescaling, and commit-time KV-cache compression to support 4-bit and NVFP4 deployment.

MC-Sparse analyzes why sparse attention can damage fidelity in long-sequence diffusion generation and introduces a training-free approach built around fine-grained KV selection and metadata reuse. The method reports substantial denoising speedups for video and 3D generation with negligible quality loss.

A new framework called Inverse Distillation Unlearning, or IDU, trains efficient one-step generators while suppressing outputs tied to a designated subset of training data. It requires only a full-data teacher and examples from the forget set.

ALoDLM introduces token-adaptive latent recurrence to diffusion language models, allowing easy positions to be committed early while difficult ones receive additional refinement. The paper reports improved average benchmark performance at 1.7B and 8B scales without giving up parallel decoding.

HC-DLM combines discrete token readouts with a persistent continuous latent state to address broken dependencies in parallel diffusion decoding. The paper reports gains over matched-size discrete and continuous diffusion baselines on Sudoku, Countdown, and LM1B.

Alibaba’s Qwen team has open-sourced Qwen-Image-2.1, a 7B visual generation model designed to balance image quality, inference efficiency, and cost. It combines text-to-image generation, transparent-image creation, and image editing in one model.

A study of 3B-parameter block-diffusion denoisers shows that state-space sequence mixers can provide exact, constant-size caches. Mamba delivers major memory, latency, and throughput gains over attention at very long contexts.

LLaDA-Image combines a from-scratch 6B Diffusion Transformer with a frozen vision-language understanding module for generation, editing, and text rendering. Its distilled Turbo variant reduces inference to 2–4 sampling steps while the project releases weights, code, and training recipes.

EditBridge introduces a diffusion bridge for ultra-high-resolution image editing. Instead of independently super-resolving a low-resolution edit, it uses the original high-resolution image to guide detail recovery while reducing cross-image attention costs.

DiffusionGemma is an experimental open-weight language model that uses discrete diffusion to generate text in parallel blocks rather than token by token. Built by fine-tuning Gemma 4, it explores whether diffusion-style decoding can deliver a better speed-capability trade-off for LLMs.

This paper studies how structured language in prompts affects diffusion-based visual generation. Rather than treating prompt length as the key factor, it shows that measurable structure in text conditioning can correlate with lower converged diffusion loss.

Explorative Modeling reframes generative training by exploring multiple candidate matches between model outputs and data, then learning from the best one. The paper argues that exploration can scale alongside parameters and data.

Chimera proposes a hybrid visual diffusion architecture designed to make high-resolution image and long-video generation more compute-efficient. It combines linear-time state tracking, intermittent global interaction, local convolutions, sparse MoE capacity, and a scaling recipe called HeteroP.

A new paper introduces Parallel Decoding Distillation, a trajectory-based distillation method designed to speed up diffusion and flow-matching models without relying heavily on adversarial or VSD-style objectives.

Sony researchers introduce Spectral Alignment, a lightweight inference-time guidance method for correcting frequency-dependent mismatch in diffusion and flow-matching models. The method calibrates intermediate predictions with a precomputed spectral prior while adding only modest overhead.

Three-Body Scattering Modeling offers a new training signal for one-step generators by learning how generated samples should move toward real data, without an adversarial critic or a prescribed diffusion path.

Mage-Flow argues that competitive image generation does not have to rely only on ever-larger parameter counts. Its 4B-scale stack combines an efficient latent tokenizer, a native-resolution diffusion transformer, and system-level CUDA optimizations.

A new paper argues that structural text template tokens in text-to-image Diffusion Transformers are not passive placeholders. They become implicit semantic registers that help preserve object identity during generation.

This paper investigates whether diffusion language models internally track denoising progress. The authors find that a latent time-like signal can be decoded from residual streams even without explicit timestep conditioning.

DiFA is a training-free inference framework for diffusion models that treats intermediate predictions as correlated observations rather than disposable integration outputs. By building a forward-aligned temporal consensus, it aims to improve stability and preserve fine details during sampling.