Diffusers 0.43.0 Released with 5 New Modular Pipelines

At a Glance
| Item | Value |
|---|---|
| Repository | huggingface/diffusers |
| Version | v0.43.0 |
| Published | 2026-11-12 |
| 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
huggingface/diffusers is a Python library (Apache-2.0) for running diffusion models for image, video, and audio generation in PyTorch. Diffusers 0.43.0 has been released, adding five new models as modular pipelines: MiniMax-H3-Image, Pika H3 Flex, P-H3-Image, Kandinsky S2V 1.0, and PAN 2.
The most impactful change for users is that minimum dependency versions are now actually enforced. Environments that do not meet torch>=2.6 or transformers>=4.41.2 will treat the corresponding packages as “unavailable". Users with older environments should check before updating.
Breaking Changes and Deprecations
| Target | Old | New |
|---|---|---|
| Dependency version check | Might pass even if below minimum version | Treated as unavailable if below minimum, and a message prompting an upgrade is displayed |
| latent upscaler / condition embedder | Imported from pipeline modules | Moved to diffusers.models. Old imports are deprecated and planned for removal in 1.0.0. Use public imports like from diffusers import LTX2TextConnectors |
| modular input defaults | Explicitly specified None handling not documented |
Explicitly specified None is preserved. Omitted inputs take the default value of the block that reads them first |
--max_grad_norm |
Disabled in some LoRA examples | Enabled |
| Component imports | Import arbitrary libraries from repository settings | Custom modules outside supported libraries require prior import on the calling side |
If an InputParam is required and has a non-None default value, it is treated as optional and a warning is issued.
Existing checkpoints are reported to be loadable even after the move. The change affects code using old import paths and environments with older torch/transformers.
--max_grad_norm was previously ineffective in train_text_to_image_lora.py and train_lcm_distill_lora_sdxl.py. Even with the default value of 1.0, training results will differ from before on runs with large gradients.
Custom modular blocks using remote code have their reliability verified before requirement processing.
Key Changes
Addition of New Models
MiniMax-H3-Image generates images from text. It also supports editing via reference images and uses a fixed 6-step schedule. It is loaded with ModularPipeline.from_pretrained(...) followed by a call to load_components(...). Pika H3 Flex builds on MiniMax-H3 to generate video and stereo audio in 6 steps. Documentation also covers how to share loaded MiniMax-H3 components between both pipelines.
P-H3-Image (Pruna AI) supports instruction-based editing with up to 8 input images and runs in 8 steps. Kandinsky S2V 1.0 generates character animation videos from reference images and audio in 5-second segments. It features Lite and Pro checkpoints, with the distilled version using PiflowScheduler defaulting to 16 steps.
PAN 2 is a 29B generative simulator. It supports text-to-video and image-to-video, utilizing context parallelism and three types of caches (FirstBlockCache, MagCache, TaylorSeer). Safety checks are enabled by default and require the separate pan2-guardrail package. All of these models are newly added, with no reconversion tasks required.
Attention for AMD RDNA4
The new aiter_sage_hub backend enables SageAttention kernels from the Hub on RDNA4 gfx1201 GPUs. Unsupported input shapes fall back to standard SDPA.
Tensor Parallelism for MiniMax-H3
Transformers can be split across multiple devices, including large modulation projections. Replicated parameter dtypes are now also preserved when loading with tensor parallelism.
Attention Backend Fixes
ModelMixin.set_attention_backend() no longer overwrites the backend for the entire process. The issue where a backend selected for one model affected unrelated models has been fixed. Additionally, FlashAttention-2 works with new Hub kernels, and Hub kernels are now downloaded when selecting the backend. This is relevant for users running multiple models in the same process.
LoRA and Checkpoint Saving
Broadcasting component weight dictionaries across multiple adapters has been fixed. Kohya-format text-encoder LoRAs preserve per-module alpha values. save_pretrained() no longer deletes shards of other variants when replacing split checkpoints with a single file.
Modular Caching and Training Examples
Stateful cache hooks are reset on every call, including calls that raise exceptions. In training examples, the FLUX quantized embedding preparation script has been updated to support the latest Transformers. An issue where an expired deprecation block caused training to fail when --revision was specified has also been resolved.
Supported Models and Hardware
In this release, the following models are available as modular pipelines:
- MiniMax-H3-Image: Supports text-to-image generation and editing using one or more reference images.
- Pika H3 Flex: Generates video and stereo audio from text, first/last frames, and multimodal references. Also supports loading LoRA adapters.
- P-H3-Image: Supports text-to-image generation and instruction-based editing with up to 8 input images, including an upsampler to detail prompts.
- Kandinsky S2V 1.0: Generates videos including full-body character movement from reference images and audio, with Lite and Pro checkpoints.
- PAN 2: Runs a 29B generative simulator for text-to-video and image-to-video, supporting context parallelism and three cache types. Separate reasoning models are out of scope for integration and are linked from the model documentation.
For hardware, SageAttention via the aiter_sage_hub backend is now available on AMD RDNA4 (gfx1201) GPUs, and MiniMax-H3 supports tensor parallelism.
Memory requirements are reported to be documented in model-specific documentation, such as for Kandinsky S2V. However, numeric VRAM figures for each model are not provided in the current materials.
How to Get It
The README recommends installing via PyPI or Conda within a virtual environment. The materials only provide recommendations without specific commands, so please check the v0.43.0 Release Page for update procedures.
Before updating, please verify the following:
- Whether
torch>=2.6andtransformers>=4.41.2are satisfied. If not, the respective packages will be treated as unavailable. - Whether latent upscjalers or condition embedders are being imported from old paths.
- Whether the separate
pan2-guardrailpackage is installed if using PAN 2.
The basic loading pattern follows this example from the README:
from diffusers import DiffusionPipeline
import torch
pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", dtype=torch.float16)
pipeline.to("cuda")
The five new models are loaded using ModularPipeline.from_pretrained(...) followed by load_components(...). Refer to each model’s documentation for details.
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 |
|---|---|---|
| v0.42.1 | 2026-10-27 | GitHub |
Related Articles
- Diffusers v0.41.0 Released with Qwen-Image 2.1 and Optimizations
- Diffusers 0.42.0 Released: FLUX.3, Fibo 2.0 and More
What to Read Next
- Follow this tool → Diffusers overview and release history (97 releases tracked)
- Other image, video and audio → ComfyUI / InvokeAI / whisper.cpp

