pypi / diffusers
diffusers
What changed in diffusers 0.37 to 0.43, from its official release notes, upgrade guides and advisories. Sources checked 2026-11-12.
21 breaking5 security4 deprecations131 new APIs and options4 performance12 behavior and defaults1 tooling47 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initdiffusers 0.4320 changes
Breaking changes 5
Checks now enforce supported dependency minimums and treat an installed package below its minimum as unavailable, with an upgrade message.
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.
Latent upscalers and condition embedders move under diffusers.models; existing checkpoints remain loadable.
Gradient clipping now takes effect in the affected LoRA training examples, changing runs whose gradients exceed the clipping threshold.
Explicitly passing None preserves that value; omitted inputs receive the default of the first block that reads them.
Security 1
Adds modular text-to-video and image-to-video workflows for scene simulation, with distributed inference, caching, and default safety checks.
Deprecations 1
Old pipeline-module imports for relocated latent upscalers and condition embedders are deprecated for removal in 1.0.0; use public imports instead.
New APIs and options 6
Adds a Hub-hosted SageAttention kernel for RDNA4 gfx1201 GPUs, with native SDPA fallback for unsupported input shapes.
Adds modular speech-driven video generation from a reference image and speech recording, with Lite and Pro checkpoints.
Adds a modular image-generation and editing integration supporting reference images, prompt enhancement, and component offloading.
Supports sharding the transformer across devices, including large modulation projections; replicated parameters retain their intended dtype during tensor-parallel loading.
Adds modular text-to-image generation and instruction-based editing with up to eight input images and prompt upsampling.
Adds modular video generation with synchronized stereo audio, multimodal references, and LoRA adapter loading.
Behavior and default changes 1
An InputParam declared both required and with a non-None default is treated as optional and emits a warning.
Fixes and other changes 6
Selecting a backend on one model no longer changes unrelated models; attention modules check requirements and download Hub kernels when selecting a backend.
The integration works with newer Hub kernel versions.
Text-encoder LoRAs retain their per-module alpha values.
Component-weight dictionaries now broadcast correctly across multiple adapters.
Stateful cache hooks reset after each pipeline call, including calls that raise an exception, so later generations start with fresh cache state.
Saving preserves shards belonging to other variants and removes obsolete indexes when replacing a sharded checkpoint with a single-file checkpoint.
diffusers 0.4225 changes
Breaking changes 4
Component library names are resolved differently across DiffusionPipeline, AutoModel, and ModularPipeline loading paths.
The supported huggingface_hub version range is now >=1.32.0,<3.0.
Qwen-Image now applies
.alphascaling to lora_A/lora_B checkpoints.Drops Python 3.10 support; Python 3.11 or newer is required to install this release.
Deprecations 1
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.
New APIs and options 14
Adds encoding and decoding for 48 kHz audio.
Adds Bria Fibo 2.0 image generation from structured JSON captions and image editing with up to five reference images.
Adds a pipeline for Kandinsky 6 video super-resolution.
Adds Kandinsky 6 video generation with synchronized audio from text or a reference image.
Adds loading support for LyCORIS Kronecker-product adapters for Z-Image and FLUX.2, including Klein.
Adds MagCache support for FLUX.2.
Adds FLUX.3 Image Klein support through modular pipelines, including text-to-image generation, editing, prompt upsampling, and reference images.
Adds LTX-2 motion control that sets the positional time base independently of playback frame rate.
Adds a scheduler for Kandinsky 6 distilled models.
Adds modular pipeline support for Qwen-Image 2.1.
Adds a separate package for preparing RFD4-Proteína model inputs and writing CIF files.
Adds camera-controlled video generation from a starting image and text prompt, with optional refinement using the separate LTX-2 refiner.
Adds tensor parallelism support for Qwen-Image 2.1 and Intel XPU, alongside TorchTPU integration.
Adds TorchTPU integration covering eager execution, torch.compile, and tensor parallelism.
Fixes and other changes 6
Loading reports a clearer error when the declared class is not found in the resolved module.
Text encoders now correctly report active adapters.
Preserves explicit timesteps.
Preserves the three-value return when no mask is supplied.
Z-Image now honors the LoRA scale passed through joint_attention_kwargs.
TorchAO versions older than 0.15 no longer break model imports; TorchAO quantization still requires version 0.15 or newer.
diffusers 0.4115 changes
Breaking changes 3
Removes the deprecated alias; use Lumina2Pipeline instead.
Removes the deprecated alias; use LuminaPipeline instead.
Changes the default for overwrite_modular_index to True.
Deprecations 1
ONNX support is deprecated; use Optimum for ONNX Runtime workflows.
New APIs and options 6
Adds SeaCache support and mixed W8A8/W8A16 denoising for ModelOpt FP8 checkpoints.
Adds fixed sampling sigmas configuration on QwenImage21Pipeline.
Adds single-file loader support for Krea 2.
Adds pipelines for generation with keyframe slots and composable spatial and temporal refinement.
Adds single-file loader support for MiniMax-H3.
Adds modular blocks for Wan 2.2 VACE.
Performance 2
Loading from existing local paths avoids unnecessary weight downloads.
Each rank reads its own slice of sharded weights directly, reducing the memory needed during loading.
Fixes and other changes 3
Fixes callback tensor inputs that were never bound in the denoising loop.
Fixes component device placement for split-device pipelines.
Fixes a cache crash when the feature shape changes between steps.
diffusers 0.4027 changes
Breaking changes 3
Skills are installed through the CLI with skills list and skills add instead of through the Makefile.
The image latent is no longer injected into visual_cond channels.
Mixed-rank LoRAs without alpha keys now mirror ranks into alphas, using the intended scale instead of an arbitrary, key-order-dependent scale.
Security 1
Fixes path traversal and arbitrary out-of-directory file reads through sharded checkpoint index files.
New APIs and options 9
Adds LoRA support for ACE-Step.
Adds img2img blocks for Anima.
Enables prompt enhancement through a separate off-the-shelf google/gemma-4-E2B-it checkpoint.
Records per-block input defaults when sibling ConditionalPipelineBlocks declare different defaults; each branch resolves its own default when it runs.
Adds multi-image instruction-guided editing that accepts 1–5 reference images and composes their elements into a new image.
Adds a second video decoder over the same latent space, so latents are interchangeable between decoders.
Adds audio-to-audio generation for Stable Audio 3.
Adds inpainting for Stable Audio 3.
Adds text-to-audio generation for high-quality stereo audio at 44.1 kHz.
Behavior and default changes 2
Device deduction now supports pipelines whose components are split across devices.
When a loaded LTX-2.5 pipeline has a duration_head, num_frames is auto-predicted by default.
Tooling 1
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.
Fixes and other changes 11
Fixes a device mismatch in AuraFlow model parallelism.
Fixes a VAE dtype mismatch when reusing an AuraFlow pipeline.
Fixes K normalization in the Cosmos 3 Edge generator.
Fixes image preprocessing for Cosmos 3.
Fixes LoRA hot-swapping recompilation when different_shapes_for_compilation is used.
Fixes adaptive stopping for batched DiffusionGemma runs.
Fixes decoding with auto-offload for Helios.
Fixes automatic discovery of local LoRA weights in offline mode.
An adapter is removed from _merged_adapters only after it has been unfused from all components.
Fixes the rectified-flow time scheduler.
Aligns snapshot_download with the latest huggingface_hub.
diffusers 0.3951 changes
Breaking changes 1
Raises the minimum supported PyTorch version to 2.6.
New APIs and options 37
Adds modular pipeline support for Anima.
Adds Any-Step video diffusion pipelines for bidirectional and FAR causal generation.
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.
Adds an automatic pipeline class for text-to-audio.
Integrates AutoRound quantization.
Adds official support for AWS Neuron devices, including Trainium and Inferentia, with
torch.compilecompatibility.Adds AWS Neuron as an officially supported device for Trainium and Inferentia.
Adds Cosmos 3 support, including world generation, physical reasoning, video-to-video generation, action-conditioned generation, and a sound encoder.
Adds Cosmos 3, including action generation, video-to-video generation, and a sound encoder.
Adds the DiffusionGemma pipeline and schedulers.
Adds a block-diffusion encoder-decoder language model pipeline and a scheduler for committing confident generated tokens.
Adds DreamLite text-to-image and image-edit pipelines.
Adds text-to-image and image-editing pipelines, including a distilled mobile pipeline for on-device, low-latency generation.
Adds
from_single_fileloading support.Adds Ideogram 4 support.
Adds a flow-matching text-to-image pipeline with structured prompt upsampling and LoRA loading support.
Adds structured prompt upsampling to Ideogram 4.
Adds Ideogram 4 LoRA training support.
Adds a mixin for loading LoRAs with Ideogram 4.
Adds support for JoyAI-JoyImage-Edit.
Adds image editing, including spatial editing capabilities for moving and rotating objects and controlling the camera.
Adds a LoRA DreamBooth trainer for Krea 2.
Adds support for loading non-diffusers Krea 2 LoRAs.
Adds support for loading non-Diffusers LoRA weights.
Adds Krea 2 text-to-image support for base and distilled checkpoints, alongside a LoRA DreamBooth trainer.
Adds the Krea 2 text-to-image pipeline.
Adds LTX-2.X IC LoRA and HDR pipelines.
Adds an LTX-2 pipeline for HDR video generation.
Adds an LTX-2 pipeline for in-context LoRA conditioning.
Adds the Motif-Video model and pipelines.
Adds support for loading a pipeline from a flat, transformer-style repository.
Adds pixel-space text-to-image generation that denoises raw RGB images directly without a VAE.
Adds
from_single_fileloading support for SkyReelsV2 and ChronoEdit transformers.Adds modular pipeline support for Stable Diffusion 3.
Adds
_dequantizesupport to the TorchAO quantizer.Adds safetensors support to the TorchAO backend.
Pipelines can now be loaded from transformer-style flat repositories.
Performance 1
Eliminates GPU synchronization overhead and CPU-to-GPU transfers across the pipeline.
Behavior and default changes 5
Enables BitsAndBytes quantization on MPS.
Allows bucket reshuffling when using DreamBooth caches.
Enables LoRA loading on the modular pipeline.
Updates the IP-Adapter integration.
Adds compatibility with AWS Neuron devices.
Fixes and other changes 7
Fixes crashes involving guidance embeddings, prompt embeddings, tensor images, and multiple images.
Adds LoRA support and aligns the pipeline with the official Cosmos repository.
Adds LoRA support for Cosmos Predict 2.5 and aligns the pipeline with the official Cosmos repository.
Fixes the ignored generator behavior.
Enables true classifier-free guidance when using precomputed negative embeddings.
Fixes true CFG with precomputed negative embeddings and tightens input validation.
Fixes GGUF handling with modules_to_not_convert and keep_in_fp32_modules.
diffusers 0.3827 changes
Breaking changes 2
Validates that eta is in the range [0, 1].
Removes the str option from the TorchAO quantization configuration.
Security 3
Diffusers has a
trust_remote_codebypass viacustom_pipelineand local custom componentsDiffusers 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
Diffusers: TOCTOU Trust Remote Code Bypass
New APIs and options 16
Adds a text-to-music pipeline that generates variable-length stereo audio at 48 kHz from prompts and optional lyrics.
Adds a pipeline for the 8B-parameter ERNIE-Image generation model.
Adds LoRA support for Ernie-Image.
Adds a modular pipeline for Ernie-Image.
Adds Flash Attention 4 as an attention backend.
Adds FlashPack loading support.
Adds a pipeline for FLUX.2 Klein inpainting.
Adds a small decoder model intended to make Flux.2 decoding faster.
Adds a modular pipeline for HunyuanVideo 1.5.
Adds a pipeline for discrete diffusion language models that generate text through block-wise iterative refinement.
Adds a text-to-audio diffusion pipeline from Meituan LongCat.
Adds a modular pipeline for LTX Video.
Adds modular support for LTX-2 and Hunyuan 1.5.
Adds the NucleusMoE-Image pipeline for image generation with a sparse mixture-of-experts architecture.
Adds support for profiling pipelines in Diffusers.
Adds ring_anything as a context-parallelism backend, with support for arbitrary key/value sequence lengths.
Performance 1
Removes compile bottlenecks from the ZImage pipeline.
Fixes and other changes 5
Fixes the shape mismatch when using precomputed prompt embeddings with num_images_per_prompt.
Fixes group offloading when using TorchAO.
Fixes image-to-video processing by preprocessing the image at pixel resolution.
Fixes an IndexError in the HunyuanVideo image-to-video pipeline.
Fixes a device mismatch when using offloading with the UniPC scheduler.
diffusers 0.3760 changes
Breaking changes 3
Removes k-diffusion pipelines.
Python 3.8 support is sunset.
Uses T5Tokenizer instead of MT5Tokenizer for compatibility with Transformers v5.0+.
Deprecations 1
Flax schedulers are now flagged as deprecated.
New APIs and options 43
Adds a decorator to simplify model definitions.
A helper applies LoRA scaling to remove boilerplate.
Model configurations can register auto_map.
Adds Representation Autoencoders as an alternative to traditional VAEs for latent-space image generation.
Adds structured image editing with JSON inputs, source images, and native masking.
Adds ChromaInpaintPipeline.
Adds an inference pipeline, scheduler, and checkpoint conversion for Cosmos Predict2.5.
Adds inference pipelines for Cosmos Transfer2.5 general tasks and autoregressive inference.
Adds inference for Cosmos-Predict2.5, a model for simulating and predicting future world states.
Adds conditional world generation with multimodal controls including edges, blurred video, segmentation, and depth.
Adds pipeline-level support for the cpu device_map.
Device maps support device types for use with offloading.
Pipeline from_pretrained gains disable_mmap.
Adds a Fibo edit pipeline.
Adds a compact pipeline that combines image generation and editing.
Adds disable_mmap support to pipeline from_pretrained loading.
Adds image generation using a hybrid autoregressive and diffusion decoder architecture.
GlmImagePipeline adds batch support and allows prompt and prior_token_ids to be provided together.
Adds video generation pipelines for Helios-14B.
Adds minute-scale video generation with the Helios model.
Adds a Laplace scheduler for DDPM.
Adds support for LongCat-Image.
Adds single-file support for LTX 2.
Adds LTX 2.0 video pipelines.
Adds an LTX2 condition pipeline.
Adds audio-conditioned text-to-video generation with synced audio, full and distilled inference, and two-stage spatial sampling.
Adds MagCache as a caching method.
Adds a new guider implementation.
Compose reusable blocks to build custom diffusion workflows, complementing the existing DiffusionPipeline class.
Modular pipelines gain support for explicit workflows and custom pipeline blocks.
Custom blocks can be validated for requirements, and modular pipeline weights can be saved to the Hub.
Adds Qwen Image Layered support.
Adds image decomposition into independently editable RGBA layers.
Implements a RAE autoencoder.
Adds MultiControlNet support for SD3 inpainting.
Adds TaylorSeer as a caching method.
Adds Ulysses Anything Attention as a context-parallelism backend.
Adds Unified Sequence Parallel attention as a context-parallelism backend.
Supports custom sigmas in UniPCMultistepScheduler.
Adds an inpaint pipeline for Z-Image.
Adds an implementation of z-image-omni-base.
Adds Z-Image-Turbo ControlNet support.
Adds the full-capacity Z-Image foundation model pipeline for image generation.
Behavior and default changes 4
Context parallelism is enabled for kernels-based attention backends.
AutoModel can use from_config with custom code.
LTX 2 encode_video accepts more input types.
Adds batch support for GLM-Image.
Fixes and other changes 9
Fixes LoRA loading for Flux Klein.
Fixes loading ModularPipelines when their modular_model_index.json contains AutoModel type hints.
Fixes an unguarded torchvision import in Cosmos Predict 2.5.
Fixes video frame extraction.
Fixes Flux2 img2img prediction.
Allows prompt and prior_token_ids to be provided simultaneously.
Fixes LTX-2 inference when num_videos_per_prompt is greater than one and classifier-free guidance is enabled.
Fixes handling of txt_seq_lens.
Diffusers is fully functional with Transformers V5.