# What changed in diffusers

> 225 changes in diffusers 0.37 to 0.43 from its official release notes, upgrade guides and advisories: 21 breaking, 4 deprecations, 5 security, 131 new APIs and options. Each one cited. Sources checked 2026-11-12. HTML: https://unstale.sh/pypi/diffusers/

To see which of these apply to a repository, run `npx @unstale/cli init` in it.

## diffusers 0.43

### Breaking changes

- Checks now enforce supported dependency minimums and treat an installed package below its minimum as unavailable, with an upgrade message. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Repository configuration no longer triggers arbitrary component-library imports; callers must import unsupported custom component modules themselves, and custom-block loading checks remote-code trust before processing requirements. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Latent upscalers and condition embedders move under diffusers.models; existing checkpoints remain loadable. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Gradient clipping now takes effect in the affected LoRA training examples, changing runs whose gradients exceed the clipping threshold. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Explicitly passing None preserves that value; omitted inputs receive the default of the first block that reads them. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)

### Security

- Adds modular text-to-video and image-to-video workflows for scene simulation, with distributed inference, caching, and default safety checks. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)

### Deprecations

- Old pipeline-module imports for relocated latent upscalers and condition embedders are deprecated for removal in 1.0.0; use public imports instead. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)

### New APIs and options

- Adds a Hub-hosted SageAttention kernel for RDNA4 gfx1201 GPUs, with native SDPA fallback for unsupported input shapes. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Adds modular speech-driven video generation from a reference image and speech recording, with Lite and Pro checkpoints. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Adds a modular image-generation and editing integration supporting reference images, prompt enhancement, and component offloading. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Supports sharding the transformer across devices, including large modulation projections; replicated parameters retain their intended dtype during tensor-parallel loading. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Adds modular text-to-image generation and instruction-based editing with up to eight input images and prompt upsampling. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Adds modular video generation with synchronized stereo audio, multimodal references, and LoRA adapter loading. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)

### Behavior and default changes

- An InputParam declared both required and with a non-None default is treated as optional and emits a warning. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)

### Fixes and other changes

- Selecting a backend on one model no longer changes unrelated models; attention modules check requirements and download Hub kernels when selecting a backend. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- The integration works with newer Hub kernel versions. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Text-encoder LoRAs retain their per-module alpha values. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Component-weight dictionaries now broadcast correctly across multiple adapters. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Stateful cache hooks reset after each pipeline call, including calls that raise an exception, so later generations start with fresh cache state. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)
- Saving preserves shards belonging to other variants and removes obsolete indexes when replacing a sharded checkpoint with a single-file checkpoint. (0.43.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.43.0)

## diffusers 0.42

### Breaking changes

- Component library names are resolved differently across DiffusionPipeline, AutoModel, and ModularPipeline loading paths. (0.42.1) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.1)
- The supported huggingface_hub version range is now >=1.32.0,<3.0. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Qwen-Image now applies `.alpha` scaling to lora_A/lora_B checkpoints. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Drops Python 3.10 support; Python 3.11 or newer is required to install this release. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)

### Deprecations

- FLUX, Krea 2, and LTX-Video core denoise groups now unpack latents; custom workflows should pass unpacked latents to decode blocks, and packed-latent fallback is deprecated for removal in 0.47.0. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)

### New APIs and options

- Adds encoding and decoding for 48 kHz audio. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds Bria Fibo 2.0 image generation from structured JSON captions and image editing with up to five reference images. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds a pipeline for Kandinsky 6 video super-resolution. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds Kandinsky 6 video generation with synchronized audio from text or a reference image. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds loading support for LyCORIS Kronecker-product adapters for Z-Image and FLUX.2, including Klein. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds MagCache support for FLUX.2. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds FLUX.3 Image Klein support through modular pipelines, including text-to-image generation, editing, prompt upsampling, and reference images. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds LTX-2 motion control that sets the positional time base independently of playback frame rate. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds a scheduler for Kandinsky 6 distilled models. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds modular pipeline support for Qwen-Image 2.1. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds a separate package for preparing RFD4-Proteína model inputs and writing CIF files. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds camera-controlled video generation from a starting image and text prompt, with optional refinement using the separate LTX-2 refiner. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds tensor parallelism support for Qwen-Image 2.1 and Intel XPU, alongside TorchTPU integration. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Adds TorchTPU integration covering eager execution, torch.compile, and tensor parallelism. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)

### Fixes and other changes

- Loading reports a clearer error when the declared class is not found in the resolved module. (0.42.1) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.1)
- Text encoders now correctly report active adapters. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Preserves explicit timesteps. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Preserves the three-value return when no mask is supplied. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- Z-Image now honors the LoRA scale passed through joint_attention_kwargs. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)
- TorchAO versions older than 0.15 no longer break model imports; TorchAO quantization still requires version 0.15 or newer. (0.42.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.42.0)

## diffusers 0.41

### Breaking changes

- Removes the deprecated alias; use Lumina2Pipeline instead. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Removes the deprecated alias; use LuminaPipeline instead. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Changes the default for overwrite_modular_index to True. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)

### Deprecations

- ONNX support is deprecated; use Optimum for ONNX Runtime workflows. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)

### New APIs and options

- Adds SeaCache support and mixed W8A8/W8A16 denoising for ModelOpt FP8 checkpoints. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Adds fixed sampling sigmas configuration on QwenImage21Pipeline. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Adds single-file loader support for Krea 2. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Adds pipelines for generation with keyframe slots and composable spatial and temporal refinement. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Adds single-file loader support for MiniMax-H3. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Adds modular blocks for Wan 2.2 VACE. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)

### Performance

- Loading from existing local paths avoids unnecessary weight downloads. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Each rank reads its own slice of sharded weights directly, reducing the memory needed during loading. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)

### Fixes and other changes

- Fixes callback tensor inputs that were never bound in the denoising loop. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Fixes component device placement for split-device pipelines. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)
- Fixes a cache crash when the feature shape changes between steps. (0.41.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.41.0)

## diffusers 0.40

### Breaking changes

- Skills are installed through the CLI with skills list and skills add instead of through the Makefile. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- The image latent is no longer injected into visual_cond channels. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Mixed-rank LoRAs without alpha keys now mirror ranks into alphas, using the intended scale instead of an arbitrary, key-order-dependent scale. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)

### Security

- Fixes path traversal and arbitrary out-of-directory file reads through sharded checkpoint index files. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)

### New APIs and options

- Adds LoRA support for ACE-Step. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Adds img2img blocks for Anima. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Enables prompt enhancement through a separate off-the-shelf google/gemma-4-E2B-it checkpoint. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Records per-block input defaults when sibling ConditionalPipelineBlocks declare different defaults; each branch resolves its own default when it runs. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Adds multi-image instruction-guided editing that accepts 1–5 reference images and composes their elements into a new image. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Adds a second video decoder over the same latent space, so latents are interchangeable between decoders. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Adds audio-to-audio generation for Stable Audio 3. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Adds inpainting for Stable Audio 3. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Adds text-to-audio generation for high-quality stereo audio at 44.1 kHz. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)

### Behavior and default changes

- Device deduction now supports pipelines whose components are split across devices. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- When a loaded LTX-2.5 pipeline has a duration_head, num_frames is auto-predicted by default. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)

### Tooling

- The CLI now writes modular_model_index.json for saved custom blocks, supports workflow for Modular pipelines, works with auto CPU offload, and saves multimodal and batched-video output. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)

### Fixes and other changes

- Fixes a device mismatch in AuraFlow model parallelism. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes a VAE dtype mismatch when reusing an AuraFlow pipeline. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes K normalization in the Cosmos 3 Edge generator. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes image preprocessing for Cosmos 3. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes LoRA hot-swapping recompilation when different_shapes_for_compilation is used. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes adaptive stopping for batched DiffusionGemma runs. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes decoding with auto-offload for Helios. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes automatic discovery of local LoRA weights in offline mode. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- An adapter is removed from _merged_adapters only after it has been unfused from all components. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Fixes the rectified-flow time scheduler. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)
- Aligns snapshot_download with the latest huggingface_hub. (0.40.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.40.0)

## diffusers 0.39

### Breaking changes

- Raises the minimum supported PyTorch version to 2.6. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)

### New APIs and options

- Adds modular pipeline support for Anima. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds Any-Step video diffusion pipelines for bidirectional and FAR causal generation. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds bidirectional and FAR causal video diffusion pipelines that adapt to arbitrary inference budgets and support text-to-video, image-to-video, and video-to-video. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds an automatic pipeline class for text-to-audio. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Integrates AutoRound quantization. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds official support for AWS Neuron devices, including Trainium and Inferentia, with `torch.compile` compatibility. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds AWS Neuron as an officially supported device for Trainium and Inferentia. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds Cosmos 3 support, including world generation, physical reasoning, video-to-video generation, action-conditioned generation, and a sound encoder. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds Cosmos 3, including action generation, video-to-video generation, and a sound encoder. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds the DiffusionGemma pipeline and schedulers. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds a block-diffusion encoder-decoder language model pipeline and a scheduler for committing confident generated tokens. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds DreamLite text-to-image and image-edit pipelines. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds text-to-image and image-editing pipelines, including a distilled mobile pipeline for on-device, low-latency generation. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds `from_single_file` loading support. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds Ideogram 4 support. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds a flow-matching text-to-image pipeline with structured prompt upsampling and LoRA loading support. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds structured prompt upsampling to Ideogram 4. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds Ideogram 4 LoRA training support. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds a mixin for loading LoRAs with Ideogram 4. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds support for JoyAI-JoyImage-Edit. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds image editing, including spatial editing capabilities for moving and rotating objects and controlling the camera. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds a LoRA DreamBooth trainer for Krea 2. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds support for loading non-diffusers Krea 2 LoRAs. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds support for loading non-Diffusers LoRA weights. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds Krea 2 text-to-image support for base and distilled checkpoints, alongside a LoRA DreamBooth trainer. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds the Krea 2 text-to-image pipeline. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds LTX-2.X IC LoRA and HDR pipelines. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds an LTX-2 pipeline for HDR video generation. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds an LTX-2 pipeline for in-context LoRA conditioning. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds the Motif-Video model and pipelines. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds support for loading a pipeline from a flat, transformer-style repository. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds pixel-space text-to-image generation that denoises raw RGB images directly without a VAE. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds `from_single_file` loading support for SkyReelsV2 and ChronoEdit transformers. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds modular pipeline support for Stable Diffusion 3. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds `_dequantize` support to the TorchAO quantizer. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds safetensors support to the TorchAO backend. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Pipelines can now be loaded from transformer-style flat repositories. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)

### Performance

- Eliminates GPU synchronization overhead and CPU-to-GPU transfers across the pipeline. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)

### Behavior and default changes

- Enables BitsAndBytes quantization on MPS. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Allows bucket reshuffling when using DreamBooth caches. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Enables LoRA loading on the modular pipeline. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Updates the IP-Adapter integration. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds compatibility with AWS Neuron devices. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)

### Fixes and other changes

- Fixes crashes involving guidance embeddings, prompt embeddings, tensor images, and multiple images. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds LoRA support and aligns the pipeline with the official Cosmos repository. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Adds LoRA support for Cosmos Predict 2.5 and aligns the pipeline with the official Cosmos repository. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Fixes the ignored generator behavior. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Enables true classifier-free guidance when using precomputed negative embeddings. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Fixes true CFG with precomputed negative embeddings and tightens input validation. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)
- Fixes GGUF handling with modules_to_not_convert and keep_in_fp32_modules. (0.39.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.39.0)

## diffusers 0.38

### Breaking changes

- Validates that eta is in the range [0, 1]. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Removes the str option from the TorchAO quantization configuration. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)

### Security

- Diffusers has a `trust_remote_code` bypass via `custom_pipeline` and local custom components (0.38.0) — [security advisory](https://github.com/advisories/GHSA-98H9-4798-4Q5V)
- Diffusers is the a library for pretrained diffusion models. Prior to 0.38.0, diffusers 0.37.0 allows remote code execution without the trust_remote_code=True safeguard when loading pipelines from Hugging Face Hub repositories. The _resolve_custom_pipeline_and_cls function in pipeline_loading_utils.py performs string interpolation on the (0.38.0) — [security advisory](https://github.com/advisories/GHSA-J7W6-VPVQ-J3GM), [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Diffusers: TOCTOU Trust Remote Code Bypass (0.38.0) — [security advisory](https://github.com/advisories/GHSA-7WX4-6VFF-V64P)

### New APIs and options

- Adds a text-to-music pipeline that generates variable-length stereo audio at 48 kHz from prompts and optional lyrics. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a pipeline for the 8B-parameter ERNIE-Image generation model. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds LoRA support for Ernie-Image. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a modular pipeline for Ernie-Image. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds Flash Attention 4 as an attention backend. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds FlashPack loading support. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a pipeline for FLUX.2 Klein inpainting. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a small decoder model intended to make Flux.2 decoding faster. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a modular pipeline for HunyuanVideo 1.5. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a pipeline for discrete diffusion language models that generate text through block-wise iterative refinement. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a text-to-audio diffusion pipeline from Meituan LongCat. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds a modular pipeline for LTX Video. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds modular support for LTX-2 and Hunyuan 1.5. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds the NucleusMoE-Image pipeline for image generation with a sparse mixture-of-experts architecture. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds support for profiling pipelines in Diffusers. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Adds ring_anything as a context-parallelism backend, with support for arbitrary key/value sequence lengths. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)

### Performance

- Removes compile bottlenecks from the ZImage pipeline. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)

### Fixes and other changes

- Fixes the shape mismatch when using precomputed prompt embeddings with num_images_per_prompt. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Fixes group offloading when using TorchAO. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Fixes image-to-video processing by preprocessing the image at pixel resolution. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Fixes an IndexError in the HunyuanVideo image-to-video pipeline. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)
- Fixes a device mismatch when using offloading with the UniPC scheduler. (0.38.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.38.0)

## diffusers 0.37

### Breaking changes

- Removes k-diffusion pipelines. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Python 3.8 support is sunset. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Uses T5Tokenizer instead of MT5Tokenizer for compatibility with Transformers v5.0+. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)

### Deprecations

- Flax schedulers are now flagged as deprecated. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)

### New APIs and options

- Adds a decorator to simplify model definitions. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- A helper applies LoRA scaling to remove boilerplate. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Model configurations can register auto_map. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds Representation Autoencoders as an alternative to traditional VAEs for latent-space image generation. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds structured image editing with JSON inputs, source images, and native masking. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds ChromaInpaintPipeline. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds an inference pipeline, scheduler, and checkpoint conversion for Cosmos Predict2.5. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds inference pipelines for Cosmos Transfer2.5 general tasks and autoregressive inference. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds inference for Cosmos-Predict2.5, a model for simulating and predicting future world states. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds conditional world generation with multimodal controls including edges, blurred video, segmentation, and depth. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds pipeline-level support for the cpu device_map. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Device maps support device types for use with offloading. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Pipeline from_pretrained gains disable_mmap. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds a Fibo edit pipeline. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds a compact pipeline that combines image generation and editing. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds disable_mmap support to pipeline from_pretrained loading. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds image generation using a hybrid autoregressive and diffusion decoder architecture. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- GlmImagePipeline adds batch support and allows prompt and prior_token_ids to be provided together. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds video generation pipelines for Helios-14B. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds minute-scale video generation with the Helios model. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds a Laplace scheduler for DDPM. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds support for LongCat-Image. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds single-file support for LTX 2. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds LTX 2.0 video pipelines. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds an LTX2 condition pipeline. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds audio-conditioned text-to-video generation with synced audio, full and distilled inference, and two-stage spatial sampling. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds MagCache as a caching method. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds a new guider implementation. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Compose reusable blocks to build custom diffusion workflows, complementing the existing DiffusionPipeline class. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Modular pipelines gain support for explicit workflows and custom pipeline blocks. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Custom blocks can be validated for requirements, and modular pipeline weights can be saved to the Hub. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds Qwen Image Layered support. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds image decomposition into independently editable RGBA layers. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Implements a RAE autoencoder. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds MultiControlNet support for SD3 inpainting. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds TaylorSeer as a caching method. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds Ulysses Anything Attention as a context-parallelism backend. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds Unified Sequence Parallel attention as a context-parallelism backend. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Supports custom sigmas in UniPCMultistepScheduler. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds an inpaint pipeline for Z-Image. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds an implementation of z-image-omni-base. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds Z-Image-Turbo ControlNet support. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds the full-capacity Z-Image foundation model pipeline for image generation. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)

### Behavior and default changes

- Context parallelism is enabled for kernels-based attention backends. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- AutoModel can use from_config with custom code. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- LTX 2 encode_video accepts more input types. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Adds batch support for GLM-Image. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)

### Fixes and other changes

- Fixes LoRA loading for Flux Klein. (0.37.1) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.1)
- Fixes loading ModularPipelines when their modular_model_index.json contains AutoModel type hints. (0.37.1) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.1)
- Fixes an unguarded torchvision import in Cosmos Predict 2.5. (0.37.1) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.1)
- Fixes video frame extraction. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Fixes Flux2 img2img prediction. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Allows prompt and prior_token_ids to be provided simultaneously. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Fixes LTX-2 inference when num_videos_per_prompt is greater than one and classifier-free guidance is enabled. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Fixes handling of txt_seq_lens. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)
- Diffusers is fully functional with Transformers V5. (0.37.0) — [release notes](https://github.com/huggingface/diffusers/releases/tag/v0.37.0)