Diffusers 0.42.0 Adds FLUX.3 Image Klein, Kandinsky 6 and SANA-WM Pipelines and Drops Python 3.10
Hugging Face's Diffusers 0.42.0 adds FLUX.3 Image Klein, Bria Fibo 2.0, Kandinsky 6, SANA-WM and RFD4-Proteína, and now requires Python 3.11 or newer.
Overview
Version 0.42.0 of the Hugging Face Diffusers library was published on October 23, 2026, according to the release page. The release notes say it brings FLUX.3 Image Klein, Bria Fibo 2.0, Kandinsky 6, SANA-WM and RFD4-Proteína to Diffusers, alongside new adapter support, hardware integrations and modular pipelines. Python 3.11 or newer is now required.
New Model Integrations
The release notes describe each addition:
- FLUX.3 Image Klein from Black Forest Labs supports text-to-image generation and editing with one or more reference images. Base and step-distilled checkpoints are available through
ModularPipeline, with an integrated prompt upsampler, single-file transformer loading, and BFL FP8r and NVFP4 checkpoints. - Bria Fibo 2.0 generates images from structured JSON captions and edits with up to five reference images through
BriaFibo2Pipeline. The edit-turbo checkpoint runs in four steps without classifier-free guidance. - Kandinsky 6 generates video with synchronized audio from text or a reference image through
Kandinsky6TI2VAPipeline, with a separateKandinsky6SRPipelinefor video super-resolution. - SANA-WM generates video from a starting image, a text prompt and a camera trajectory, which can be given as explicit poses or a WASD/IJKL action string.
- 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.
FLUX.3 Image Klein Details
The Diffusers documentation describes FLUX 3 Image Klein as a 5.6B image generation model from Black Forest Labs, with text conditioning from Qwen3-VL. The pipeline runs a prompt upsampler first on every request. The upsampler is a separate 9B model from Black Forest Labs that expands the prompt into a full scene caption before the text encoder.
Per the same documentation, the pipeline carries recommended sampling settings of 50 steps, guidance scale 3.5 and scheduler shift 6.93, and generates 1024 x 1024 images by default. The step-distilled checkpoint generates in 16 steps without classifier-free guidance.
On memory, the documentation says loading everything in bf16 takes about 39 GB of weights, and a pipeline built without the upsampler step loads about 20 GB. For GPUs with less memory, it says to enable component offloading.
Kandinsky 6 Details
The Kandinsky 6 documentation describes a family of video generation models from Kandinsky Lab. A single multimodal diffusion transformer denoises video and audio latents together. Distilled checkpoints ship with the few-step PiflowScheduler and must be run with guidance_scale=1.0. The super-resolution pipeline upscales by 2, 4 or 2.25.
Compatibility Changes
The release notes list several changes for existing users:
- Python 3.10 support is dropped.
- The supported
huggingface_hubrange is now>=1.32.0,<3.0. - FLUX, Krea 2 and LTX-Video now unpack latents in their core denoise groups, so custom modular workflows should pass unpacked latents to their decode blocks. The fallback for packed latents is deprecated for removal in 0.47.0.
- Having a TorchAO version older than 0.15 installed no longer breaks model imports, though TorchAO quantization still requires version 0.15 or newer.
Other items in the release include TorchTPU integration covering eager execution, torch.compile and tensor parallelism, Intel XPU support for tensor parallelism, LoKr adapter loading for Z-Image and FLUX.2, and MagCache support for FLUX.2.
What We Don’t Know
The sources reviewed are project-published release notes and documentation. They do not include independent benchmarks or user reports for the new pipelines, so real-world quality and speed on the new models are not covered here.