diffusion

A UCLA-led paper showing that long-range planning can be learned from purely random exploration using a conditional energy-based model estimating temporal log-density ratios, with no action or reward labels, achieving multiscale cognitive-map-like navigation from states, images, and egocentric observation.

NVIDIA introduces PixelUMM, an encoder-free unified multimodal model that reads and writes raw pixels with a single decoder-only Transformer (Qwen3-8B backbone), removing VAEs and ViT encoders from image and video understanding and generation. Code and model checkpoints are released alongside detailed ablations on patch artifacts and decoder head design.

EMA Lightning is a tiny, fast Turkish text-to-speech model with 8.6M total parameters (~34 MB) that claims the lowest WER (0.92%) on Freya-TR-Eval, beating ElevenLabs v4 and Gemini 3.8, with streaming generation and fully offline operation under Apache 2.0.

UniEvo-VL 提出了一种用于多模态模型的在线自蒸馏训练方案，让单一模型同时充当教师与学生（教师条件于自生成的批评信息），以提升其图像生成能力，在 GenEval 上从 0.747 提升至 0.808。

SILSA is a topology-aware 3D generation framework that replaces expensive voxel tokens with compact sliding-window slice latents and single-stage rectified-flow generation, using persistence-diagram/Betti-based slice-level topology supervision. It improves structural fidelity (PSNR +8.7%, Betti error -9.2%) while cutting tokens by 70%, memory by 40.4%, and inference time by 58.5%.

GroundAnything 提出一个 4B 参数的视觉定位基础模型，通过块状去噪协调并行解码与精确定位，在 30 个 grounding 基准上取得同类模型新 SOTA（72.42%），并借助自投机解码实现 4.51 倍加速。

UniEvo-VL introduces an on-policy self-distillation training recipe where a single multimodal model acts as both teacher (conditioned on its own privileged self-critiques) and student, improving image generation on Qwen-Image-2512 with gains from 0.747 to 0.808 on GenEval without external supervision.

The paper proposes Future-Aware Recall (FAR), a framework that learns episodic memory recall for world models via future-aware predictive supervision and adaptive multi-cue scoring, automatically determining which retrieval cues (time, pose, vision, audio) to trust.

Prism 是一个动态稀疏注意力框架，用于原生训练 2K 分辨率的联合视频-音频生成模型，通过时空宏区块划分与内容自适应的块形状分配，在实现 2.5 倍训练加速的同时超越全注意力的生成质量，由复旦大学与腾讯混元联合发布。

A tweet highlights how price cuts in AI models like Opus 5.5 and GPT-6 Sol/Luna are reducing costs and enabling broader AI use-cases through the Jevons paradox, accelerating economic diffusion.

SGLang-Diffusion with VDN-H3 from MiniMax enables fast and scalable video generation, achieving 14.4s of 768p video in just 9.0s on 8× B200 GPUs for faster-than-real-time inference.

DiffIE introduces a diffusion-based method for open information extraction that uses stochastic reverse-diffusion to generate multiple candidate triplets, achieving state-of-the-art performance on benchmarks like CaRB and BenchIE.

Viggle-Animate is an AI model that replaces characters in videos using a single repainted frame, achieving efficient and accurate motion transfer without complex intermediate representations.

Evoke is an interactive world model with external persistent memory and a redesigned long-horizon teacher, enabling open-ended video generation with bounded context and low latency. It achieves state-of-the-art performance on WBench while staying competitive on VBench-Long and VBench-2.0.

Anima-2.9B is a fine-tuned and layer-expanded anime/illustration diffusion model built on circlestone-labs/Anima, with new layers trained on 1.7M samples and a July 2026 knowledge cutoff. It is currently in preview and requires a custom ComfyUI node for use.

This paper proposes Second-Order Drifting Models, which augment drifting generative models with artificial velocity variables to achieve accelerated second-order dynamics in Fourier space, mitigating spectral stiffness while preserving one-step inference. The method is evaluated on synthetic distribution matching, sequential data generation, and robotic control, showing improved convergence and competitive performance.

This paper investigates when classifier-free guidance (CFG) is actually necessary in masked diffusion language models, showing that guidance dependence is prompt-specific and can often be removed without losing constraint satisfaction, leading to a defined 'commitment horizon'.

ReRound is a post-training quantization method that uses a conditional diffusion model to guide rounding of near-midpoint weights, improving small LLM accuracy at 3-bit and 4-bit quantization without inference overhead.

The paper introduces Adversarial Fréchet Distance (AdvFD), which adds a learnable adversarial feature space to static Fréchet losses to improve generator post-training, with real-feature whitening to stabilize optimization.

An annotated from-scratch reimplementation of Google's DiffusionGemma, a 26B open-weight state diffusion language model, explaining its architecture, sampling procedure, and design choices.