This release comes packed with new image generation and editing pipelines, a new video pipeline, new training scripts, quality-of-life improvements, and much more. Read the rest of the release notes fully to not miss out on the fun stuff.

New pipelines 🧨

We welcomed new pipelines in this release:

• Wan 2.2
• Flux-Kontext
• Qwen-Image
• Qwen-Image-Edit

Wan 2.2 📹

This update to Wan provides significant improvements in video fidelity, prompt adherence, and style. Please check out the official doc to learn more.

Flux-Kontext 🎇

Flux-Kontext is a 12-billion-parameter rectified flow transformer capable of editing images based on text instructions. Please check out the official doc to learn more about it.

Qwen-Image 🌅

After a successful run of delivering language models and vision-language models, the Qwen team is back with an image generation model, which is Apache-2.0 licensed! It achieves significant advances in complex text rendering and precise image editing. To learn more about this powerful model, refer to our docs.

Thanks to @naykun for contributing both Qwen-Image and Qwen-Image-Edit via this PR and this PR.

New training scripts 🎛️

Make these newly added models your own with our training scripts:

• Kontext trainer
• Qwen-Image trainer

Single-file modeling implementations

Following the 🤗 Transformers’ philosophy of single-file modeling implementations, we have started implementing modeling code in single and self-contained files. The Flux Transformer code is one example of this.

Attention refactor

We have massively refactored how we do attention in the models. This allows us to provide support for different attention backends (such as PyTorch native scaled_dot_product_attention, Flash Attention 3, SAGE attention, etc.) in the library seamlessly.

Having attention supported this way also allows us to integrate different parallelization mechanisms, which we’re actively working on. Follow this PR if you’re interested.

Users shouldn’t be affected at all by these changes. Please open an issue if you face any problems.

Regional compilation

Regional compilation trims cold-start latency by only compiling the small and frequently-repeated block(s) of a model - typically a transformer layer - and enables reusing compiled artifacts for every subsequent occurrence. For many diffusion architectures, this delivers the same runtime speedups as full-graph compilation and reduces compile time by 8–10x. Refer to this doc to learn more.

Thanks to @anijain2305 for contributing this feature in this PR.

We have also authored a number of posts that center around the use of torch.compile. You can check them out at the links below:

• Presenting Flux Fast: Making Flux go brrr on H100s
• torch.compile and Diffusers: A Hands-On Guide to Peak Performance
• Fast LoRA inference for Flux with Diffusers and PEFT

Faster pipeline loading ⚡️

Users can now load pipelines directly on an accelerator device leading to significantly faster load times. This particularly becomes evident when loading large pipelines like Wan and Qwen-Image.

from diffusers import DiffusionPipeline
import torch 

ckpt_id = "Qwen/Qwen-Image"
pipe = DiffusionPipeline.from_pretrained(
-    ckpt_id, torch_dtype=torch.bfloat16
- ).to("cuda")
+    ckpt_id, torch_dtype=torch.bfloat16, device_map="cuda"
+ )				

You can speed up loading even more by enabling parallelized loading of state dict shards. This is particularly helpful when you’re working with large models like Wan and Qwen-Image, where the model state dicts are typically sharded across multiple files.

import os
os.environ["HF_ENABLE_PARALLEL_LOADING"] = "yes"

# rest of the loading code
....

Better GGUF integration

@Isotr0py contributed support for native GGUF CUDA kernels in this PR. This should provide an approximately 10% improvement in inference speed.

We have also worked on a tool for converting regular checkpoints to GGUF, letting the community easily share their GGUF checkpoints. Learn more here.

We now support loading of Diffusers format GGUF checkpoints.

You can learn more about all of this in our GGUF official docs.

Modular Diffusers (Experimental)

Modular Diffusers is a system for building diffusion pipelines pipelines with individual pipeline blocks. It is highly customisable, with blocks that can be mixed and matched to adapt to or create a pipeline for a specific workflow or multiple workflows.

The API is currently in active development and is being released as an experimental feature. Learn more in our docs.

All commits

• [tests] skip instead of returning. by @sayakpaul in #11793
• adjust to get CI test cases passed on XPU by @kaixuanliu in #11759
• fix deprecation in lora after 0.34.0 release by @sayakpaul in #11802
• [chore] post release v0.34.0 by @sayakpaul in #11800
• Follow up for Group Offload to Disk by @DN6 in #11760
• [rfc][compile] compile method for DiffusionPipeline by @anijain2305 in #11705
• [tests] add a test on torch compile for varied resolutions by @sayakpaul in #11776
• adjust tolerance criteria for test_float16_inference in unit test by @kaixuanliu in #11809
• Flux Kontext by @a-r-r-o-w in #11812
• Kontext training by @sayakpaul in #11813
• Kontext fixes by @a-r-r-o-w in #11815
• remove syncs before denoising in Kontext by @sayakpaul in #11818
• [CI] disable onnx, mps, flax from the CI by @sayakpaul in #11803
• TorchAO compile + offloading tests by @a-r-r-o-w in #11697
• Support dynamically loading/unloading loras with group offloading by @a-r-r-o-w in #11804
• [lora] fix: lora unloading behvaiour by @sayakpaul in #11822
• [lora]feat: use exclude modules to loraconfig. by @sayakpaul in #11806
• ENH: Improve speed of function expanding LoRA scales by @BenjaminBossan in #11834
• Remove print statement in SCM Scheduler by @a-r-r-o-w in #11836
• [tests] add test for hotswapping + compilation on resolution changes by @sayakpaul in #11825
• reset deterministic in tearDownClass by @jiqing-feng in #11785
• [tests] Fix failing float16 cuda tests by @a-r-r-o-w in #11835
• [single file] Cosmos by @a-r-r-o-w in #11801
• [docs] fix single_file example. by @sayakpaul in #11847
• Use real-valued instead of complex tensors in Wan2.1 RoPE by @mjkvaak-amd in #11649
• [docs] Batch generation by @stevhliu in #11841
• [docs] Deprecated pipelines by @stevhliu in #11838
• fix norm not training in train_control_lora_flux.py by @Luo-Yihang in #11832
• [From Single File] support from_single_file method for WanVACE3DTransformer by @J4BEZ in #11807
• [lora] tests for exclude_modules with Wan VACE by @sayakpaul in #11843
• update: FluxKontextInpaintPipeline support by @vuongminh1907 in #11820
• [Flux Kontext] Support Fal Kontext LoRA by @linoytsaban in #11823
• [docs] Add a note of _keep_in_fp32_modules by @a-r-r-o-w in #11851
• [benchmarks] overhaul benchmarks by @sayakpaul in #11565
• FIX set_lora_device when target layers differ by @BenjaminBossan in #11844

Significant community contributions

The following contributors have made significant changes to the library over the last release:

• @vuongminh1907
  • update: FluxKontextInpaintPipeline support (#11820)
• @Net-Mist
  • feat: add multiple input image support in Flux Kontext (#11880)
• @tolgacangoz
  • Add SkyReels V2: Infinite-Length Film Generative Model (#11518)
• @naykun
  • Qwen-Image (#12055)
  • fix(qwen-image): update vae license (#12063)
  • Qwen Image Edit Support (#12164)
• @Trgtuan10
  • Add QwenImage Inpainting and Img2Img pipeline (#12117)
• @SamYuan1990
  • try to use deepseek with an agent to auto i18n to zh (#12032)