AI Briefing#diffusion-models
The latest AI and developer news about #diffusion-models, with the original source and a short summary.
The latest AI and developer news about #diffusion-models, with the original source and a short summary.
A new script and container allow users to fine-tune TRELLIS.2 with LoRA on custom 3D assets, significantly improving geometric accuracy for specific domains like footwear.
The model couples a Qwen3.5-9B language model with a 29B MMDiT diffusion backbone to predict video outcomes of actions.
The model compresses MiniMax-H3 to 13.7B parameters for 4-8 step generation, running on consumer RTX 5090 GPUs in 92 seconds for 5.2-second clips.
The method removes inter-round dependencies by drafting successor blocks during verification, achieving higher average speedups than DFlash across seven benchmarks.
The arXiv paper (2605.29713) offers a derivation-oriented guide connecting PCA, VAEs, diffusion models, and GANs for researchers and students.
A Jane Street intern project demonstrated that flow matching provides stable generation of synthetic order book events where DDPM diverged, though fully continuous diffusion struggles with the discrete nature of market microstructure.
Lunara uses fewer than 10B active parameters and achieves the highest mean scores in blinded human evaluation across aesthetic quality, emotional resonance, and content integrity.
Apple researchers introduced Normalizing Trajectory Models (NTM), a framework that matches strong baselines in four sampling steps while retaining exact likelihood over the generative trajectory.
Linum AI proposes Pyramid-JiT, a VAE-free pixel-space text-to-image model that achieves an FD-DINOv2 score of 27.9, approaching the real-image floor of 25.4, while requiring 11.3x fewer training samples than Linum v2.
Apple research shows confidence-based sampling in discrete diffusion models selects dependent token groups, causing a 29x deviation from the training distribution despite perfect per-sample metrics.
Apple researchers have unveiled DSAS, a dynamic scaling Activation Steering technique designed to reduce unnecessary performance degradation.
A study has been released that transforms language models into continuous spaces and distills them to perform code generation in a single step.
A method that converts auto-regressive models into diffusion models to make them **more than 2x** faster.
SD3.5-Flash is a lightweight distillation framework that produces high-quality images in just a few steps.
ARC post-training generates 12 seconds of 44.1kHz stereo audio in 75ms on an H100.