Diffusers v0.41.0 Released with Qwen-Image 2.1 and Optimizations

Diffusers v0.41.0 Released with Qwen-Image 2.1 and Optimizations

At a Glance

Item Value
Repository huggingface/diffusers
Version v0.41.0
Published 2026-10-06
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

Latest version “v0.41.0" of “huggingface/diffusers" (Python, Apache-2.0 license), a state-of-the-art diffusion models library for generating images, videos, and audio in PyTorch, has been released.

The most significant change in this version is the integration of “Qwen-Image 2.1", the latest model combining image generation and editing. This enables text-to-image generation, image editing, and even LoRA training directly within Diffusers. Additionally, starting with this release, a release policy similar to “Transformers" has been adopted, aligning minor releases with new model integrations.

Breaking Changes & Deprecations

This version deprecates ONNX support and completely removes older APIs that were previously deprecated. Please check if your code requires the following changes before updating.

Old Feature / API Migration Path / Recommended Method
ONNX support (deprecated) Use Optimum for ONNX Runtime workflows.
LuminaText2ImgPipeline (removed) Use LuminaPipeline.
Lumina2Text2ImgPipeline (removed) Use Lumina2Pipeline.

Affected Users

This affects developers building workflows with ONNX or those calling older aliases of Lumina models (such as LuminaText2ImgPipeline) in their code. Since these old code paths have been removed, code modifications are required before updating.

Key Changes

Integration of Qwen-Image 2.1 Pipeline

“Qwen-Image 2.1", which integrates image generation and editing into a single model, is now available. The visual generation component of this model has a 7B parameter scale and supports generation using multiple reference images as well as native RGBA output for transparent images. Furthermore, LoRA training is also supported. This affects users who want to perform advanced image generation and editing using Qwen-Image 2.1.

Memory Reduction via Tensor-Parallel Loading

When loading sharded checkpoints during tensor-parallel processing, each rank now directly reads its own slice. This significantly reduces the amount of memory required when loading models. This benefits users loading large models in parallel across multi-GPU environments.

Addition of LTX-2.5 DFR Pipeline

Support for the new LTX-2.5 DFR pipeline has been added. This enables advanced generation combining keyframe slot-based generation and spatial/temporal refinement. This affects users looking to leverage video generation workflows with LTX-2.5.

Cosmos 3 Enhancements

SeaCache support has been added to the Cosmos 3 pipeline. Additionally, mixed W8A8 and W8A16 denoising is now supported for ModelOpt FP8 checkpoints. This affects users running Cosmos 3 models efficiently with FP8 or similar configurations.

Modularization and Local Loading Optimization

Modular blocks have been added to Wan 2.2 VACE. Improvements were also made to avoid downloading unnecessary weights when loading models from existing local paths. For developers repeatedly loading and testing models in local environments, this cuts down on redundant downloads and reduces startup times.

Expansion of Single-File Loaders

Single-file loaders for Krea 2 and MiniMax-H3 are newly supported. This allows users to easily load these models from a single file and run inference.

Supported Models and Hardware

This version includes numerous updates for new model architectures and optimizations for specific hardware environments.

Qwen-Image 2.1

Added support for “Qwen-Image 2.1", which integrates image generation and editing capabilities into a single model. The visual generation component of this model is sized at 7B (7 billion) parameters, supporting generation with multiple reference images and native RGBA output for transparent images. In Diffusers 0.41.0, text-to-image generation, image editing, LoRA training, and single file format loading (single file support) are supported.

LTX-2.5 DFR

Added support for the new LTX-2.5 DFR pipeline. Advanced generation processes combining keyframe slot-based generation and spatial/temporal refinement become available.

Cosmos 3

Caching technology “SeaCache" is newly supported. It also supports mixed W8A8 and W8A16 denoising for ModelOpt FP8 checkpoints. In addition, bug fixes related to FP8 tensor-parallel processing and issues regarding SeaCache artifact transfer when using control CFG have been addressed.

Wan 2.2 VACE

A modular blockset has been added for Wan 2.2 VACE, enabling more flexible pipeline construction.

Krea 2

Krea 2 models are newly supported with single file format loading. Additionally, processing to tag Krea 2 text encoder outputs as denoiser_input_fields has been added.

MiniMax-H3

MiniMax-H3 models are newly supported with single file format loading. For backends that do not support fp64 (double-precision floating point), logic to downcast position_ids during device transfer was added, along with fixes for VAE decoding processes.

Hardware and Execution Environment Improvements

Operational stability and optimizations for specific GPUs and accelerator environments have also been implemented.

  • Improved Ascend NPU Support: Fixed issues related to mask shapes for RMSNorm and fused-attention in Huawei Ascend NPU environments.
  • Test Optimization for Non-CUDA Devices: Modified environment checks to automatically skip certain modeling tests containing CUDA-only operations on devices where CUDA is unavailable.
  • Strengthened Tensor Parallelism and Distributed Environments: Introduced a mechanism where each rank directly reads/writes its own shard during tensor-parallel checkpoint loading and saving, improving memory efficiency. Computation stream synchronization was also added prior to offloading to disk. In distributed training environments, the behavior of the learning rate (LR) scheduler when num_train_epochs is specified was fixed.

How to Get It

🤗 Diffusers is recommended to be installed from PyPI or Conda within a virtual environment. Please check the official installation guide and release page for specific installation steps and the latest dependencies.

After installation, you can easily try loading a model and generating images as follows. The following code demonstrates loading a model from over 30,000 checkpoints published on the Hub using the from_pretrained method to generate an image from text:

from diffusers import DiffusionPipeline
import torch

pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", dtype=torch.float16)
pipeline.to("cuda")
pipeline("An image of a squirrel in Picasso style").images[0]

Alternatively, if you want to customize models and schedulers in greater detail to build your own diffusion system, please refer to the following code example:

from diffusers import DDPMScheduler, UNet2DModel
from PIL import Image
import torch

scheduler = DDPMScheduler.from_pretrained("google/ddpm-cat-256")
model = UNet2DModel.from_pretrained("google/ddpm-cat-256").to("cuda")
scheduler.set_timesteps(50)

sample_size = model.config.sample_size
noise = torch.randn((1, 3, sample_size, sample_size), device="cuda")
input = noise

for t in scheduler.timesteps:
    with torch.no_grad():
        noisy_residual = model(input, t).sample
        prev_noisy_sample = scheduler.step(noisy_residual, t, input).prev_sample
        input = prev_noisy_sample

image = (input / 2 + 0.5).clamp(0, 1)
image = image.cpu().permute(0, 2, 3, 1).numpy()[0]
image = Image.fromarray((image * 255).round().astype("uint8"))
image

Releases Since Our Last Article

Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.

Version Released Release notes
v0.40.0 2026-08-20 GitHub
v0.39.0 2026-07-03 GitHub
v0.38.0 2026-05-01 GitHub
v0.37.1 (Fixes for AutoModel type hints in Modular Pipelines and Flux Klein LoRA loading ) 2026-03-25 GitHub
v0.37.0 2026-03-06 GitHub
v0.36.0 2025-12-08 GitHub
v0.35.2 (🐞 fixes for transformers models, imports, ) 2025-10-15 GitHub
v0.35.1 2025-08-20 GitHub
v0.35.0 2025-08-19 GitHub
v0.34.0 2025-06-25 GitHub
v0.33.1 2025-04-10 GitHub
v0.33.0 2025-04-09 GitHub
v0.32.2 2025-01-16 GitHub
v0.32.1 2024-12-25 GitHub
v0.32.0 2024-12-24 GitHub
v0.31.0 2024-10-22 GitHub
v0.30.3 2024-09-17 GitHub
v0.30.2 2024-08-31 GitHub
v0.30.1 2024-08-24 GitHub
v0.30.0 2024-08-07 GitHub
v0.29.2 2024-06-27 GitHub
v0.29.1 2024-06-21 GitHub
v0.29.0 2024-06-13 GitHub
v0.28.2 2024-06-05 GitHub
v0.28.1 2024-06-04 GitHub
v0.28.0 2024-05-27 GitHub
v0.27.2 2024-03-20 GitHub
v0.27.1 2024-03-19 GitHub
v0.27.0 2024-03-15 GitHub
v0.26.3 2024-02-13 GitHub
v0.26.2 2024-02-06 GitHub
v0.26.1 2024-02-02 GitHub
v0.26.0 2024-02-01 GitHub
v0.25.1 (Patch release) 2024-01-18 GitHub
v0.25.0 2023-12-27 GitHub
v0.24.0 2023-11-30 GitHub
v0.23.1 ([Patch release] Make sure we install correct PEFT version) 2023-11-16 GitHub
v0.23.0 2023-11-10 GitHub
v0.22.3 2023-11-08 GitHub
v0.22.2 2023-11-08 GitHub
v0.22.1 (Patch Release: Fix community vs. hub pipelines revision) 2023-11-06 GitHub
v0.22.0 2023-11-06 GitHub
v0.21.4 (Patch Release: Fix Lora fusing/unfusing) 2023-09-30 GitHub
v0.21.3 (Patch Release: Fix LoRA attention processor for xformers.) 2023-09-28 GitHub
v0.21.2 (Patch Release: CPU offloading + Lora load/Text inv load & Multi Adapter) 2023-09-19 GitHub
v0.21.1 2023-09-14 GitHub
v0.21.0 2023-09-14 GitHub
v0.20.2 2023-09-01 GitHub
v0.20.1 (Patch Release: Fix torch.compile() support for ControlNets) 2023-08-28 GitHub
v0.20.0 2023-08-17 GitHub
v0.19.3 (Patch release: Fix incorrect filenaming) 2023-07-30 GitHub
v0.19.2 (Patch Release: Support for SDXL Kohya-style LoRAs, Fix batched inference SDXL Img2Img, Improve watermarker) 2023-07-29 GitHub
v0.19.1 (Patch Release: Fix torch compile and local_files_only) 2023-07-28 GitHub
v0.19.0 2023-07-27 GitHub
v0.18.2 2023-07-12 GitHub
v0.18.1 (Patch Release for Stable Diffusion XL 0.9) 2023-07-08 GitHub
v0.18.0 (Shap-E, Consistency Models, Video2Video) 2023-07-07 GitHub
v0.17.1 2023-06-12 GitHub
v0.17.0 2023-06-09 GitHub
v0.16.1 2023-04-28 GitHub
v0.16.0 2023-04-27 GitHub
v0.15.1 2023-04-18 GitHub
v0.15.0 2023-04-13 GitHub
v0.14.0 (ControlNet, 8K VAE decoding) 2023-03-04 GitHub
v0.13.1 2023-02-20 GitHub
v0.13.0 (Controllable Generation: Pix2Pix0, Attend and Excite, SEGA, SAG, …) 2023-02-20 GitHub
v0.12.1 2023-01-27 GitHub
v0.12.0 (Instruct-Pix2Pix, DiT, LoRA) 2023-01-26 GitHub
v0.11.1 2022-12-21 GitHub
v0.11.0 2022-12-20 GitHub
v0.10.2 2022-12-10 GitHub
v0.10.1 2022-12-09 GitHub
v0.10.0 2022-12-09 GitHub
v0.9.0 2022-11-26 GitHub
v0.8.1 2022-11-24 GitHub
v0.8.0 2022-11-24 GitHub
v0.7.2 2022-11-06 GitHub
v0.7.1 2022-11-04 GitHub
v0.7.0 2022-11-04 GitHub
v0.6.0 2022-10-20 GitHub
v0.5.1 2022-10-14 GitHub
v0.5.0 2022-10-14 GitHub
v0.4.2 2022-10-12 GitHub
v0.4.1 2022-10-07 GitHub
v0.4.0 2022-10-07 GitHub
v0.3.0 2022-09-09 GitHub
v0.2.4 2022-08-23 GitHub
v0.2.3 2022-08-22 GitHub
v0.2.2 2022-08-17 GitHub
v0.2.1 2022-08-17 GitHub
v0.2.0 2022-08-17 GitHub
0.1.3 2022-07-28 GitHub
0.1.2 (Initial release of 🧨 Diffusers) 2022-07-21 GitHub

Related Articles

What to Read Next

Sources