diffusion

DiffusionGemma technical report released on arXiv, with ongoing work on llama.cpp pull requests to enable faster local inference on limited VRAM.

PRISM introduces a theory for designing reference processes in Schrödinger bridge models, showing that under finite computational budgets the optimal reference noise spectrum is determined by the sensor's information destruction spectrum. Experiments confirm the theory in Gaussian settings and identify where real images deviate.

Hugging Face page for the Minimax-h3-Turbo video generation model, with instructions for using it via Diffusers and Colab/Kaggle notebooks.

A researcher muses on the unstable equilibrium where language uses autoregressive models while other modalities use diffusion, and speculates that a unified multimodal architecture depends on the order each modality reveals information. He seals a bet on this idea.

DiffImaginE is a research paper proposing a diffusion-based verifier for multimodal named entity recognition, replacing deterministic imagination with conditional latent diffusion inference for more robust entity type verification.

DiffusionGemma is an experimental open-weight language model that generates text via discrete diffusion rather than token-by-token decoding, enabling exceptionally high-speed generation.

JoyAI-Video-Edit is a 16B-parameter autoregressive diffusion framework for real-time, open-ended video editing, achieving 720p editing at ~30 FPS on a single NVIDIA B200 GPU.

Hunyuan3D-Buffalo 1.0 is a unified multimodal model for 3D generation, understanding, and editing, trained on an 87M-scale 3D multimodal corpus. It combines Hunyuan3D-VLM and Hunyuan3D-DiT to achieve state-of-the-art performance on text-to-3D generation and 3D editing benchmarks.

AURORA-LM introduces a continuous-latent diffusion language model that separates decodable text representation from distribution modeling, achieving strong performance on OpenWebText and XSum while scaling to 1B parameters.

Tutorial on implementing a diffusion model based on Google's Gemma architecture from scratch.

LeapTalk proposes a novel framework that overcomes the latency-quality trade-off in talking head generation via single-step bridge distillation, enabling stable real-time generation at up to 200 FPS with reduced identity drift.

MotifRole-Diff proposes a role-aware corruption schedule for masked discrete diffusion on molecular graphs, allocating masking rates based on denoising difficulty and graph-level perturbation impact, demonstrating improved validity and reduced FCD on QM9 and MOSES benchmarks.

This paper identifies a failure mode in classifier-free guidance distillation called Negative Branch Asymmetry, where errors in the positive and negative CFG branches cancel out, and proposes Positive-Direction Matching to supervise branches separately for more robust distilled models.

Celeris-1 is a new language model using diffusion-based inference architecture, achieving near-GPT-5 level intelligence with 15x faster response times and high throughput.

Lightricks releases LTX-2.5, an open-weights world model for generating synchronized video and audio from text, image, and video inputs, with features like native multishot generation and a new diffusion video decoder.

Proposes Self Gradient Forcing (SGF), a two-pass training strategy for autoregressive video diffusion models that provides missing supervision for writing useful context memory, enabling strong long-video extrapolation even from short training windows.

Neural Drive is an action-conditioned world model for SuperTuxKart that runs entirely in the browser via WebGPU, predicting game frames from past frames and controls without a game engine.

Released a model (Stereo2Spatial) that converts stereo music tracks to spatialized binaural mixes, using flow-matching diffusion and amplitude lifting for stable training. The model and a Windows app are open-sourced under Apache 2.0.

This paper proposes that large-scale text-to-video generation can serve as a powerful pre-training paradigm for computer vision, introducing GenCeption which achieves state-of-the-art performance across diverse vision tasks with high data efficiency and emergent generalization to unseen domains.

This paper proposes using intrinsic device noise on analog neuromorphic hardware as a resource for continual learning by conditioning each weight's stochastic dynamics to avoid crossing memory-critical barriers, demonstrating non-monotonic retention improvement and validation on BrainScaleS-2 silicon.