Diffusers 0.42.0 Released: FLUX.3, Fibo 2.0 and More

At a Glance
| Item | Value |
|---|---|
| Repository | huggingface/diffusers |
| Version | v0.42.0 |
| Published | 2026-10-24 |
| License | Apache-2.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Hugging Face has released version 0.42.0 of “Diffusers", a PyTorch-based library for working with state-of-the-art diffusion models (image, video, and audio generation). This release includes support for Black Forest Labs’ latest model, “FLUX.3 Image Klein", as well as biomolecular design and advanced video generation model integrations. The most significant change for users is that Python 3.11 or higher is now required, meaning those on older environments must update their Python setup before upgrading.
Breaking Changes and Deprecations
| Item | Old Specification | New Specification |
|---|---|---|
| Python Version | 3.10 or higher | 3.11 or higher |
| huggingface_hub | < 2.0 | >= 1.32.0, < 3.0 |
| Latent format (FLUX, etc.) | Unpack inside decoder | Unpack inside denoise groups |
| TorchAO Version | (No restriction) | 0.15 or higher (when using quantization) |
Because support for Python 3.10 has ended, users must upgrade to Python 3.11 or later before installing the latest version. Additionally, for FLUX, Krea 2, and LTX-Video models, the latent variable unpacking process has been moved to the end of the denoising process. Developers building custom workflows will need to modify their code to pass the unpacked latents to the decode block. The traditional packed format is deprecated and is scheduled to be removed in version 0.47.0.
Key Changes
Support for FLUX.3 Image Klein
Black Forest Labs’ latest model, “FLUX.3 Image Klein", is now supported. It enables not only text-to-image generation but also image editing using one or more reference images, as well as generation based on reference images.
– What Changed: Base and step-distilled versions of FLUX.3 are now available through ModularPipeline. A prompt upsampler that converts short prompts into detailed captions has been integrated, and layout specification via bounding boxes is also supported.
– Who Is Affected: Users who want to perform high-quality image generation and editing using FLUX.3.
– Actions Before/After Update: Modify your implementation to load the pipeline with ModularPipeline.from_pretrained() and load weights using load_components().
Integration of Bria Fibo 2.0
“Bria Fibo 2.0" has been added, supporting image generation from structured JSON-format captions and editing using up to five reference images.
– What Changed: BriaFibo2Pipeline has been introduced, making the dev checkpoints and the “edit-turbo" checkpoints (which operate in 4 steps without Classifier-Free Guidance (CFG)) available.
– Who Is Affected: Users wanting to perform precise image editing using structured prompts or multiple images.
– Actions Before/After Update: Use the new pipeline class and prepare inputs according to the specific prompt format described in the documentation.
Audio-inclusive Video Generation with Kandinsky 6
“Kandinsky 6" has been integrated, capable of generating videos with synchronized audio from text or reference images.
– What Changed: Generation via Kandinsky6TI2VAPipeline and video super-resolution via Kandinsky6SRPipeline are now supported. PiflowScheduler has also been introduced for distilled models.
– Who Is Affected: Users looking to generate videos with audio or perform video upscaling.
– Actions Before/After Update: Check the specific requirements described in the documentation if you intend to use the super-resolution feature.
SANA-WM and Camera Control
“SANA-WM" is now supported, allowing users to generate videos by specifying images, text, and camera trajectories.
– What Changed: Camera control via action strings such as WASD/IJKL or explicit poses is now available. A two-stage workflow that enhances quality using an LTX-2 refiner post-generation is also supported.
– Who Is Affected: Users who want fine-grained control over camera work within videos.
– Actions Before/After Update: Refer to the SANA-WM documentation to check the input format for camera control.
Biomolecular Design with RFD4-Proteína
“RFD4-Proteína" has been introduced to simultaneously generate protein sequences and all-atom structures.
– What Changed: Pipelines, transformers, autoencoders, and schedulers for biomolecular design have been integrated. Molecular contexts (such as binding targets and motifs) can be specified as conditions.
– Who Is Affected: Users leveraging Diffusers for scientific computing such as drug discovery and protein engineering.
– Actions Before/After Update: You must separately install the rfd4-proteina-prep package to prepare input data and export CIF files.
Hardware Support and Optimization Expansions
TorchTPU integration and tensor parallelism support for Intel XPU have been added.
– What Changed: Eager execution, torch.compile, and tensor parallelism are now supported on TPUs. Tensor parallelism has also been made available for Qwen-Image 2.1.
– Who Is Affected: Users running TPUs, Intel XPU, or Qwen-Image 2.1 in large-scale environments.
– Actions Before/After Update: Verify your latest TorchTPU backend settings in TPU environments. Also, fixes have been applied so that model imports do not fail even in environments where TorchAO older than 0.15 is installed, though 0.15 or higher is still required for quantization features.
Supported Models and Hardware
Newly Supported Models and Pipelines
This version enables inference and utilization of several state-of-the-art models. The main added models are as follows:
- FLUX.3 Image Klein: Added support for Black Forest Labs’ latest model. Base and distilled versions can be loaded via
ModularPipeline. The ability to directly load BFL FP8r and NVFP4 checkpoints has also been integrated. - Bria Fibo 2.0: Introduced
BriaFibo2Pipeline. It supports image generation via structured JSON prompts and editing using up to five reference images. The 4-step “edit-turbo" checkpoints are also supported. - Kandinsky 6: Audio-synchronized video generation via
Kandinsky6TI2VAPipelineand video super-resolution viaKandinsky6SRPipelineare now available, supporting Pro, Lite, and distilled checkpoints. - SANA-WM: An image-to-video generation pipeline with camera control features has been added. Camera work can be controlled using key input-style action strings like WASD and IJKL, or explicit pose specifications.
- RFD4-Proteína: A model for biomolecular design has been integrated. It enables simultaneous generation of protein sequences and all-atom structures, including transformers, autoencoders, schedulers, and LoRA loading functionality.
- Qwen-Image 2.1: New modular pipelines (
QwenImage21ModularPipelineandQwenImage21TurboModularPipeline) have been added. In addition to text-to-image generation, generation using one or more reference images is also supported.
Quantization Formats and Adapter Expansions
Features related to model lightweighting and customization have also been enhanced.
- LoKr Adapters: Added the ability to load LyCORIS Kronecker-product (LoKr) adapters. These can be used with Z-Image and the FLUX.2 family, including FLUX.3 Klein.
- KVAE-Audio:
AutoencoderKLKVAEAudiohas been added to encode and decode 48 kHz audio, which can be utilized in high-quality audio generation workflows. - MagCache: Caching functionality is now supported in FLUX.2 to streamline inference efficiency.
Hardware Accelerator Optimizations
Performance improvements have been implemented across diverse computing environments.
- TorchTPU: Eager execution,
torch.compile, and tensor parallelism (TP) have been integrated for Google TPU environments, enabling fast inference and training on TPUs. - Intel XPU: Tensor parallelism is now supported on Intel GPUs (XPUs).
- AWS Trainium/Inferentia (Neuron): RoPE (Rotary Positional Embedding) optimizations for Neuron were implemented in Qwen-Image 2.1. Instead of complex number paths, real-valued cos/sin processing is applied on Neuron, achieving efficiency through frequency data caching. Block ID construction logic to avoid
repeat_interleave, which is non-executable on Neuron, has also been introduced. - Tensor Parallelism for Qwen-Image 2.1: Tensor parallelism has been implemented in
QwenImage21Transformer2DModelto split Attention layers (to_q/to_k/to_v) and SwiGLU layers along the column dimension, and output projection layers along the row dimension.
How to Get It
To use this version, first ensure that your Python environment is 3.11 or higher. Then, perform the update using the following steps.
For standard installations or upgrades, run the following command from PyPI:
pip install --upgrade diffusers
If you want to use the newly added biomolecular design model RFD4-Proteína, you must separately install the dedicated auxiliary package for input data preparation and CIF file output:
pip install rfd4-proteina-prep
Additionally, when using quantization features (TorchAO), the torchao version must be 0.15 or higher. Since older installed versions may cause model imports to fail, updating is recommended concurrently.
Related Articles
- Diffusers v0.41.0 Released with Qwen-Image 2.1 and Optimizations
- Diffusers 0.43.0 Released with 5 New Modular Pipelines
- Qwen-Image-2.1-Turbo 8-Step Fast Image Generation Model: 48GB+ VRAM
- Qwen-Image-2.1-viggle-turbo Image Generation Model: 48GB+ VRAM
What to Read Next
- Follow this tool → Diffusers overview and release history (97 releases tracked)
- Other image, video and audio → ComfyUI / InvokeAI / whisper.cpp
Sources
- huggingface/diffusers Diffusers 0.42.0: FLUX.3, Fibo 2.0, Kandinsky 6, and more
- PR #15030: ci: move to python 3.11.
- PR #14737: move unpack out of the flux, krea2 and ltx decoders into the core denoise groups
- PR #15095: Fix model imports when torchao is older than 0.15
- PR #14863: Bump huggingface_hub upper bound to <3.0 (diffusers)
- PR #14976: add modular pipeline for qwen-image 2.1
- PR #14865: [core] Tensor parallelism for Qwen-Image-2.1
- PR #14039: [TPU] TorchTPU backend integration – eager / torch.compile / tp
- PR #15013: xpu supports TP as well
- PR #14163: [LoRA] add LoKr adapter support (Z-Image, Flux2/Klein)
- PR #14967: [LTX-2] Add motion_speed to condition the positional time base
- PR #14384: [feat] add KVAE-Audio autoencoder
- PR #14756: [hooks] Support Flux2 in MagCache
- PR #15088: Add RFD4-Proteina

