AI Briefing#diffusion-model
The latest AI and developer news about #diffusion-model, with the original source and a short summary.
The latest AI and developer news about #diffusion-model, with the original source and a short summary.
Linum AI releases P-JiT, a 2B pixel-space text-to-image model trained for $21,038, which removes the VAE to achieve 4x attention context reduction.
The model enables precise pixel-level changes without unintentionally altering unrelated details, addressing a common limitation of latent-space generators.
GravityOCR, based on GLM-OCR, increased decoding speed by 1.32x without adding parameters while maintaining an accuracy score of 95.16.
Google Research and UIUC unveiled R4T, which distills RL-trained query fan-out behaviors into a 53.9M parameter diffusion model, reducing search latency by 12–20x.
IFM has released K2-Horizon-7B, which combines a diffusion adapter with a causal LLM, claiming inference speeds of up to 5200 tps and lossless performance.
The JiT-DDT architecture, adopting a pixel-space encoder-decoder structure, achieved faster convergence and improved image quality while reducing GPU usage by 3.6x compared to Linum v2.
A technique combining Diffusion draft generation and AR verification with the GLM-OCR model has been released, increasing decoding speed by 3.85x.
Continuous diffusion models, once touted as an alternative to autoregressive models, are showing signs of revival amid a recent resurgence in research.
Discrete diffusion-based open-weight LLM 'DiffusionGemma' demonstrates ultra-fast inference performance.
Tencent ARC released SCoPE, a video generation model that controls camera trajectories.
JoyAI has released a 16B model that causally edits live video streams in real time based on natural language instructions.
Apple researchers implemented 4-step text generation using a 1.7B CFM model.
Bidirectional diffusion models estimate rollout errors using forward-backward round-trip discrepancies.
DiffusionGemma adapts Gemma 4 into a discrete diffusion model, achieving 1,500 TPS on H100.
Explorative Modeling adds an exploration axis to generative models to boost efficiency.
A new study shows that regularization in neural networks smooths the score function, guiding diffusion models toward creative interpolation.
A next-generation language model, celeris-1, has been unveiled, applying Diffusion technology instead of the traditional Autoregressive approach to revolutionize speed.
POCKET-Image achieves accurate Korean text rendering even on-device through a pipeline that preserves glyph shapes.
Hugging Face Diffusers has officially integrated Nunchaku W4A4 quantization based on SVDQuant, cutting VRAM usage in half and improving inference speed.
Microsoft has unveiled Mage-Flow, an image generation and editing model that rivals large models with just 4B parameters.
Vidu S1, an interactive video model that responds in real time to users' voice instructions and generates video indefinitely, has been unveiled.
CalibAtt is a technique that speeds up video generation models by up to 1.58x by identifying unnecessary connections during attention computation.
NVIDIA and HuggingFace have unveiled a NeMo Automodel integration for fine-tuning diffusion models.
The Wan Team has released WanSong, a diffusion-based music generation model.