MOSS-Image, a 5B-parameter open-source image generation model by MOSI and OpenMOSS, surpassed 1.3 million downloads within four days of release. By excluding AI-generated images from its training data, the model effectively mitigates the 'AI look,' achieving high naturalness that sparked discussion in overseas communities. Developers have rapidly released quantized, accelerated, and LoRA variants, while validating its utility in layout control and diverse style generation.
DiT
扩散变换器
A diffusion model variant using Transformer architecture as the denoising backbone
Curated Intelligence
Sand AI's 114B-parameter audio-video model MAGI-2 Preview now runs on Moore Threads MTT S5000. Leveraging vLLM-Omni and the torchada adapter, execution path optimizations reduced denoising latency from 15.6s to 565ms per step.
Linum introduces Pyramid-JiT (P-JiT), a decoder-only pixel-space architecture that eliminates the need for a Variational Autoencoder (VAE). By predicting target images at multiple resolutions along the DiT trunk, P-JiT achieves comparable FD-DINOv2 scores to their previous latent diffusion model Linum v2 using 11.3x fewer training samples and 4.3x fewer GPU-hours, while supporting 4x higher pixel resolution.
Alibaba's Qwen team has open-sourced Qwen-Image-2.1, a lightweight 7B-parameter model that unifies text-to-image generation and image editing within a single architecture. The model natively supports transparent image generation and editing, leveraging a mixed-granularity attention mechanism to optimize inference efficiency while enhancing fidelity in typography, lighting, and multi-reference local edits.
The vLLM-Omni framework implements system-wide full-stack optimizations for the MiniMax H3 model, covering attention mechanisms, communication, and parallel VAE decoding. It further integrates FastVideo's FastH3 technique to reduce DiT forward passes from 49 to just 4. Benchmarks on eight B300 GPUs demonstrate that this approach generates a complete 10.125-second MP4 video in approximately 8.7 seconds, achieving 'real-time' performance where generation speed exceeds playback duration.
vLLM-Omni introduces Distributed Layerwise Offload (DLO), enabling large DiT video generation models exceeding single-device HBM capacity (e.g., Cosmos3-Super 64B) to run efficiently across multi-GPU/NPU clusters. By leveraging meta-device initialization, mmap weight loading, and overlapped AllGather operations, DLO reduces cold-start peak memory by 73% while maintaining a fixed double-buffer scheme for resident weights.