Diffusers 0.42.0 Adds FLUX.3 and Kandinsky 6, Requires Python 3.11

Diffusers 0.42.0: FLUX.3, Bria Fibo 2.0, Kandinsky 6 and SANA-WM #

The Diffusers 0.42.0 release notes list five model integrations, alongside new adapter support, hardware integrations and modular pipelines. The notes open with Python 3.11 or newer, and they state that Python 3.10 support is dropped. Environments still on 3.10 need an interpreter change before installing this release. Each model entry names its own pipeline class, so the code change for each model is separate.

FLUX.3 Image Klein text-to-image and editing #

FLUX.3 Image Klein, from Black Forest Labs, supports text-to-image generation and editing with one or more reference images. Both the base and step-distilled checkpoints are available through ModularPipeline. The integrated prompt upsampler turns short requests into detailed captions and accepts layout instructions with bounding boxes. A docs commit in the same release is titled "prompt upsampler required, ours or any LLM", so the FLUX.3 documentation is the place to confirm how the upsampler is expected to be used.

The notes say to load the pipeline with ModularPipeline.from_pretrained(...) and then load its weights with load_components(...). The integration also includes single-file transformer loading and BFL FP8r and NVFP4 checkpoints. The FLUX.3 documentation covers prompt upsampling, editing and memory use, so teams that adopt these checkpoints should review those sections before changing their memory settings.

Bria Fibo 2.0 generation from structured JSON captions #

Bria Fibo 2.0 generates images from structured JSON captions and edits images with up to five reference images through BriaFibo2Pipeline. The integration includes a dev checkpoint and an edit-turbo checkpoint. The notes state that edit-turbo runs in four steps without classifier-free guidance. Code that builds prompts should follow the supported prompt formats in the Fibo 2.0 documentation.

LoRA loading and DreamBooth LoRA training examples are included for text-to-image and image editing. Teams that already train LoRAs for image generation can use these examples as a starting point for Fibo 2.0. The notes do not describe how the examples differ from the existing ones, so the example code is the place to check that.

Kandinsky 6 video with synchronized audio #

Kandinsky 6 generates video with synchronized audio from text or a reference image through Kandinsky6TI2VAPipeline. A separate Kandinsky6SRPipeline provides video super-resolution. The integration supports Pro and Lite generation checkpoints, including distilled variants, and introduces PiflowScheduler for the distilled models. The notes point to the Kandinsky 6 documentation for checkpoint-specific settings and super-resolution requirements.

SANA-WM camera-controlled video #

SANA-WM generates video from a starting image, a text prompt and a camera trajectory. The camera can be controlled with explicit poses or with a WASD/IJKL action string. The output can optionally be refined with the separate LTX-2 refiner. The notes describe this as a two-stage workflow, and the SANA-WM documentation covers its steps.

RFD4-Proteína biomolecular design #

RFD4-Proteína, from the Baker Lab and NVIDIA, jointly generates protein sequences and all-atom structures while conditioning on molecular context such as binding targets, motifs or ligands. The integration includes the pipeline, transformer, autoencoder, schedulers and LoRA loading. Preparing model inputs and writing CIF files uses the separate rfd4-proteina-prep package. Teams that write CIF output will therefore need that package in addition to Diffusers.

Hardware support for TorchTPU, Intel XPU and Qwen-Image tensor parallelism #

The notes list TorchTPU integration covering eager execution, torch.compile and tensor parallelism (#14039). Tensor parallelism also gains Intel XPU support (#15013). Qwen-Image also gains new modular pipelines and tensor parallelism support in this release (#14976, #14865). Teams on TPU or Intel XPU hardware should check these entries against their deployment setup.

LoKr, motion speed and cache adapter additions #

LoKr adapters can now be loaded for Z-Image and FLUX.2, including Klein (#14163). LTX-2 gains motion_speed, which controls the positional time base independently of playback frame rate (#14967). The notes suggest a compatible adapter, such as the LTX-2.5 slow-motion control LoRA, for speeds far from the default. MagCache now supports FLUX.2 (#14756).

The release also adds AutoencoderKLKVAEAudio, which adds encoding and decoding for 48 kHz audio (#14384). Teams that process 48 kHz audio through the autoencoder can check this addition against their current audio path. The notes list these items as additions and do not describe them as changes to existing behaviour.

Adapter and pipeline fixes #

Z-Image now honors the LoRA scale in joint_attention_kwargs (#15015). Qwen-Image applies .alpha scaling to lora_A and lora_B checkpoints (#15016). Text encoders correctly report active adapters (#14961). Code that sets a LoRA scale for Z-Image through joint_attention_kwargs is the code path this fix addresses.

FlowMatchEulerDiscreteScheduler now preserves explicit timesteps (#14956). The Anima text encoder exports and standalone core denoising are fixed (#14943). InpaintProcessor.preprocess keeps its three-value return when no mask is supplied (#14481). Pipelines that use these components should be retested after upgrading.

Minimum Python version and dependency pins #

Diffusers 0.42.0 release notes: Python 3.11 or newer is now required.

The supported huggingface_hub range is now >=1.32.0,<3.0 (#14863). TorchAO versions older than 0.15 no longer break model imports, but TorchAO quantization still requires version 0.15 or newer (#15095). Environments that pin either package should check their pins against these ranges before installing.

Modular latent outputs and the packed-latent deprecation #

FLUX, Krea 2 and LTX-Video now unpack latents in their core denoise groups (#14737). Custom workflows should pass unpacked latents to their decode blocks. The fallback for packed latents is deprecated for removal in 0.47.0, so workflows that still depend on it should move to unpacked latents before that version.

Diffusers 0.41.0: Qwen-Image 2.1 and removed deprecated APIs #

The Diffusers 0.41.0 release notes introduce Qwen-Image 2.1 and state that deprecated APIs have been removed. The removal includes the LuminaText2ImgPipeline and Lumina2Text2ImgPipeline aliases. Code that imports either alias should use LuminaPipeline or Lumina2Pipeline instead.

Diffusers 0.41.0 release notes: Older integrations should review the removed code paths before upgrading.

Qwen-Image generation, editing and transparent output #

Qwen-Image unifies image generation and editing in one model. The notes say its visual generation component has 7B parameters and supports multiple reference images. The model also produces native RGBA output for transparent images. Teams that produce transparent assets can use that output directly, and the documentation covers the details.

Tensor-parallel checkpoint loading and LTX-2.5 DFR pipelines #

Tensor-parallel checkpoint loading lets each rank read its own slice of sharded weights directly (#14544). The notes say this reduces the memory needed during loading. Deployments that load sharded checkpoints across ranks are the ones affected by this change. LTX-2.5 DFR adds new pipelines that support generation with keyframe slots and composable spatial and temporal refinement (#14567).

Cosmos 3, Wan 2.2 VACE and single-file loaders #

Cosmos 3 gains SeaCache support (#14663) and mixed W8A8/W8A16 denoising for ModelOpt FP8 checkpoints (#14664). Wan 2.2 VACE gains modular blocks (#14432). Loading from existing local paths avoids unnecessary weight downloads (#14797). Single-file loaders are added for Krea 2 and MiniMax-H3 (#14914, #14839).

ONNX deprecation and the Optimum path #

ONNX support is deprecated in this release (#14658). The notes direct ONNX Runtime workflows to Optimum. Teams that depend on ONNX export should plan that migration now, because the deprecation is the step the notes announce.

Release policy, modular saving and test layout #

The notes state that minor releases are now coordinated around new model integrations and include other changes merged alongside them. Patch releases remain focused on fixes. ModularPipeline.save_pretrained now defaults overwrite_modular_index to True (#14659). Code that saves modular pipelines should confirm that this default matches its save workflow.

For contributors, the notes say the monolithic test_pipelines_common.py suite was removed to establish more modularity and transparent tests (#14669). Test-suite changes do not affect runtime code, so this entry matters mainly to people maintaining Diffusers forks.

Upgrade checklist for Diffusers #

Work through these steps in order, because the Python and dependency checks come before any code change. Python 3.11 or newer is required by the 0.42.0 release.

1. Confirm the interpreter is Python 3.11 or newer, since Python 3.10 support is dropped in 0.42.0.
2. Confirm huggingface_hub falls in the range >=1.32.0,<3.0.
3. If TorchAO is used for quantization, confirm the installed version is 0.15 or newer.
4. Replace any LuminaText2ImgPipeline or Lumina2Text2ImgPipeline imports with LuminaPipeline or Lumina2Pipeline, as 0.41.0 removes those aliases.
5. In custom modular workflows, pass unpacked latents to decode blocks for FLUX, Krea 2 and LTX-Video, since the packed-latent fallback is deprecated for removal in 0.47.0.
6. Check code that passes a LoRA scale through joint_attention_kwargs for Z-Image, since 0.42.0 honors that scale.
7. Review code that calls ModularPipeline.save_pretrained, since overwrite_modular_index now defaults to True in 0.41.0.
8. If you use ONNX Runtime workflows, plan a move to Optimum, since 0.41.0 deprecates ONNX support.
9. For each new model you plan to use, follow its documentation for the pipeline class and checkpoint settings.