# Diffusers 0.42.0 adds FLUX.3, Kandinsky 6 video, and biomolecular design tools

> **Key takeaway:** Hugging Face's Diffusers 0.42.0 introduces five major model integrations including FLUX.3 Image Klein, video generation with synchronized audio via Kandinsky 6, and protein structure prediction with RFD4-Proteína, while dropping Python 3.10 support.

**Published:** 2026-10-23T15:39:55+00:00 
**Source:** Hugging Face Diffusers 
**Category:** ai-open-models 
**Canonical URL:** https://fosswire.org/news/diffusers-0420-adds-flux3-kandinsky-6-video-and-biomolecular.html 

## Summary
Hugging Face has released Diffusers 0.42.0, a major update to its popular diffusion model library. The release integrates five new models: FLUX.3 Image Klein for text-to-image generation with reference image editing; Bria Fibo 2.0 for JSON-captioned image generation; Kandinsky 6 for text- or image-to-video with synchronized audio; SANA-WM for camera-controlled video generation from images; and RFD4-Proteína for biomolecular protein design. The update also brings TorchTPU and Intel XPU hardware support, LoKr adapter loading, and modular pipeline improvements. Python 3.11 or newer is now required, ending Python 3.10 compatibility.

## What's new in Diffusers 0.42.0
- **FLUX.3 Image Klein integration:** Black Forest Labs' model supports text-to-image generation and editing with one or more reference images, available through ModularPipeline with both base and step-distilled checkpoints. The integration includes single-file transformer loading, BFL FP8r and NVFP4 checkpoint support, and an integrated prompt upsampler that converts short requests into detailed captions with bounding box layout instructions.
- **Kandinsky 6 video generation:** The Kandinsky6TI2VAPipeline generates video with synchronized audio from text or reference images, with a separate Kandinsky6SRPipeline for video super-resolution. Pro and Lite generation checkpoints are supported including distilled variants, with a new PiflowScheduler for distilled models.
- **SANA-WM camera-controlled video:** Generates video from a starting image, text prompt, and camera trajectory controlled via explicit poses or WASD/IJKL action strings, with optional refinement through the separate LTX-2 refiner in a two-stage workflow.
- **RFD4-Proteína biomolecular design:** From the Baker Lab and NVIDIA, this integration jointly generates protein sequences and all-atom structures conditioned on molecular context such as binding targets, motifs, or ligands. Model input preparation and CIF file output require the separate rfd4-proteina-prep package.
- **Hardware and adapter expansions:** TorchTPU integration covers eager execution, torch.compile, and tensor parallelism; tensor parallelism also gains Intel XPU support. LoKr adapters enable loading LyCORIS Kronecker-product adapters for Z-Image and FLUX.2 including Klein. Python 3.11 or newer is now required.

## Why it matters
This release significantly broadens Diffusers' scope from pure image generation into video, audio, and scientific computing domains, making it a more central toolkit for AI researchers and application developers on Linux. The hardware expansion to TorchTPU and Intel XPU improves accessibility for developers without NVIDIA GPUs, while the modular pipeline architecture and standardized adapter interfaces reduce integration friction across the open-source AI model ecosystem.