v0.37.0

✨ New Features

• Introduction of Modular Diffusers, allowing pipeline construction via reusable blocks.
• New image pipeline: Z Image Omni Base, a full-capacity, undistilled transformer for high-quality generation.
• New image pipeline: Flux2 Klein, unifying generation and editing with fast inference on consumer hardware.
• New image pipeline: Qwen Image Layered, capable of decomposing images into independent RGBA layers for editability.
• New image pipeline: FIBO Edit, an 8B parameter image-to-image model using JSON inputs for structured control.
• New image pipeline: Cosmos Predict2.5, specialized for simulating and predicting future world states.
• New image pipeline: Cosmos Transfer2.5, a conditional world generation model with adaptive multimodal control.
• New image pipeline: GLM-Image, using a hybrid autoregressive + diffusion decoder architecture for high visual fidelity.
• New model: Representation Autoencoders (RAE) as an alternative to traditional VAEs.
• New video/audio pipeline: LTX-2, an audio-conditioned text-to-video generation model supporting synced audio.
• New video pipeline: Helios, a 14B parameter video generation model supporting minute-scale generation.
• Introduction of MagCache for improved caching.
• Introduction of TaylorSeer for improved caching.
• Introduction of Unified Sequence Parallel attention as a new context-parallelism (CP) backend.
• Introduction of Ulysses Anything Attention as a new context-parallelism (CP) backend.
• New Mambo-G Guidance implementation.
• New Laplace Scheduler for DDPM.
• Support for Custom Sigmas in UniPCMultistepScheduler.
• Support for MultiControlNet in SD3 Inpainting.
• Support for Context parallel in native flash attention.
• Support for NPU Ulysses Attention.
• Introduction of @apply_lora_scale decorator for simplifying model definitions.
• Introduction of pipeline-level “pu” device_map.