[LoRA] add LoKr adapter support (Z-Image, Flux2/Klein)

.github/workflows/build_documentation.yml

@@ -17,7 +17,7 @@ permissions:
 
 jobs:
   build:
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
+    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@9fc8a41d7729c424fdb2192f2f85ff63102ec9a3  # main
     with:
       commit_sha: ${{ github.sha }}
       install_libgl1: true

.github/workflows/build_pr_documentation.yml

@@ -44,7 +44,7 @@ jobs:
 
   build:
     needs: check-links
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@e60a538eea9817ab312196d0d233604b01697265  # main
+    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@9fc8a41d7729c424fdb2192f2f85ff63102ec9a3  # main
     with:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}

.github/workflows/upload_pr_documentation.yml

@@ -11,7 +11,7 @@ permissions:
 
 jobs:
   build:
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@9ad2de8582b56c017cb530c1165116d40433f1c6  # main
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@9fc8a41d7729c424fdb2192f2f85ff63102ec9a3  # main
     with:
       package_name: diffusers
     secrets:

docs/source/en/_toctree.yml

@@ -24,6 +24,8 @@
       title: Schedulers
     - local: using-diffusers/weighted_prompts
       title: Prompting
+    - local: using-diffusers/image_quality
+      title: FreeU
     - local: using-diffusers/reusing_seeds
       title: Reproducibility
     - local: using-diffusers/callback
@@ -68,24 +70,14 @@
     title: Inference
   - isExpanded: false
     sections:
-    - local: optimization/pruna
-      title: Pruna
-    - local: optimization/xformers
-      title: xFormers
-    - local: optimization/tome
-      title: Token merging
-    - local: optimization/deepcache
-      title: DeepCache
     - local: optimization/cache_dit
       title: CacheDiT
-    - local: optimization/tgate
-      title: TGATE
-    - local: optimization/xdit
-      title: xDiT
     - local: optimization/para_attn
       title: ParaAttention
-    - local: using-diffusers/image_quality
-      title: FreeU
+    - local: optimization/pruna
+      title: Pruna
+    - local: optimization/xdit
+      title: xDiT
     title: Community methods
   - isExpanded: false
     sections:
@@ -210,6 +202,8 @@
     title: Training methods
   - local: training/nemo_automodel
     title: NeMo Automodel
+  - local: training/jobs
+    title: Hugging Face Jobs
   title: Train and fine-tune
 - isExpanded: false
   sections:
@@ -375,6 +369,8 @@
         title: JoyImageEditPlusTransformer3DModel
       - local: api/models/transformer_joyimage
         title: JoyImageEditTransformer3DModel
+      - local: api/models/kandinsky6_transformer
+        title: Kandinsky 6 Transformers
       - local: api/models/krea2_transformer2d
         title: Krea2Transformer2DModel
       - local: api/models/latte_transformer3d
@@ -501,6 +497,8 @@
         title: AutoencoderSAME
       - local: api/models/consistency_decoder_vae
         title: ConsistencyDecoderVAE
+      - local: api/models/kandinsky6_vae
+        title: Kandinsky 6 VAEs
       - local: api/models/ltx2_diffusion_decoder
         title: LTX2VideoDiffusionDecoderModel
       - local: api/models/autoencoder_oobleck
@@ -725,6 +723,8 @@
         title: HunyuanVideo1.5
       - local: api/pipelines/kandinsky5_video
         title: Kandinsky 5.0 Video
+      - local: api/pipelines/kandinsky6
+        title: Kandinsky 6
       - local: api/pipelines/latte
         title: Latte
       - local: api/pipelines/ltx2
@@ -816,6 +816,8 @@
       title: LMSDiscreteScheduler
     - local: api/schedulers/minimax_h3
       title: MiniMaxH3Scheduler
+    - local: api/schedulers/piflow
+      title: PiflowScheduler
     - local: api/schedulers/pndm
       title: PNDMScheduler
     - local: api/schedulers/repaint

docs/source/en/api/models/kandinsky6_transformer.md

@@ -0,0 +1,53 @@
+<!--Copyright 2026 The Kandinsky Team and The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Kandinsky 6 Transformers
+
+Kandinsky 6 uses a multimodal diffusion transformer that denoises video and audio latents together for
+text/image-to-video-and-audio generation, and a text-free diffusion transformer for video super-resolution.
+
+## Kandinsky6Transformer3DModel
+
+The multimodal transformer used by [`Kandinsky6TI2VAPipeline`].
+
+```python
+import torch
+from diffusers import Kandinsky6Transformer3DModel
+
+transformer = Kandinsky6Transformer3DModel.from_pretrained(
+    "kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16
+)
+```
+
+[[autodoc]] Kandinsky6Transformer3DModel
+  - all
+  - forward
+
+## Kandinsky6SRTransformer3DModel
+
+The text-free transformer used by [`Kandinsky6SRPipeline`] to refine one tile of the upscaled video at a time.
+
+```python
+import torch
+from diffusers import Kandinsky6SRTransformer3DModel
+
+transformer = Kandinsky6SRTransformer3DModel.from_pretrained(
+    "kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16
+)
+# The transformer always runs NABLA sparse attention (`nabla_threshold`, 0.8 by default) on the `flex` backend.
+# Compile it, otherwise flex falls back to an eager implementation that needs far more memory at video resolutions.
+transformer.compile_repeated_blocks(fullgraph=True)
+```
+
+[[autodoc]] Kandinsky6SRTransformer3DModel
+  - all
+  - forward

docs/source/en/api/models/kandinsky6_vae.md

@@ -0,0 +1,66 @@
+<!--Copyright 2026 The Kandinsky Team and The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Kandinsky 6 VAEs
+
+Kandinsky 6 uses a causal 3D K-VAE for video super-resolution and the MMAudio mel-spectrogram VAE, paired with a
+separate BigVGAN [`MMAudioVocoder`], for synchronized audio generation.
+
+## Kandinsky6SRVAE
+
+The causal 3D K-VAE used by [`Kandinsky6SRPipeline`]. It processes arbitrarily long videos in bounded-memory
+segments while reproducing the exact output of a single, non-segmented pass.
+
+```python
+import torch
+from diffusers import Kandinsky6SRVAE
+
+vae = Kandinsky6SRVAE.from_pretrained(
+    "kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers", subfolder="vae", torch_dtype=torch.bfloat16
+)
+```
+
+[[autodoc]] Kandinsky6SRVAE
+  - encode
+  - decode
+  - all
+
+## MMAudioVAE
+
+The mel-spectrogram VAE used by [`Kandinsky6TI2VAPipeline`] when `sample_audio=True`. Its `decode` output is a mel
+spectrogram; pass it through [`MMAudioVocoder`] to get a waveform.
+
+The reference implementation can be found at [hkchengrex/MMAudio](https://github.com/hkchengrex/MMAudio) (MIT
+license).
+
+```python
+import torch
+from diffusers import MMAudioVAE
+
+audio_vae = MMAudioVAE.from_pretrained(
+    "kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers", subfolder="audio_vae", torch_dtype=torch.bfloat16
+)
+```
+
+[[autodoc]] MMAudioVAE
+  - encode
+  - decode
+  - all
+
+## MMAudioVocoder
+
+Adapted from the BigVGAN-v2 vocoder MMAudio bundles, itself from
+[NVIDIA/BigVGAN](https://github.com/NVIDIA/BigVGAN) (MIT license), with the anti-aliased Snake activations of
+[alias-free-torch](https://github.com/junjun3518/alias-free-torch) (Apache License 2.0).
+
+[[autodoc]] MMAudioVocoder
+  - forward

docs/source/en/api/pipelines/kandinsky.md

@@ -712,7 +712,7 @@ make_image_grid([img.resize((512, 512)), image.resize((512, 512))], rows=1, cols
 
 Kandinsky is unique because it requires a prior pipeline to generate the mappings, and a second pipeline to decode the latents into an image. Optimization efforts should be focused on the second pipeline because that is where the bulk of the computation is done. Here are some tips to improve Kandinsky during inference.
 
-1. Enable [xFormers](../../optimization/xformers) if you're using PyTorch < 2.0:
+1. Enable [xFormers](../../optimization/attention_backends) if you're using PyTorch < 2.0:
 
 ```diff
   from diffusers import DiffusionPipeline

docs/source/en/api/pipelines/kandinsky6.md

@@ -0,0 +1,150 @@
+<!--Copyright 2026 The Kandinsky Team and The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Kandinsky 6
+
+Kandinsky 6 is a family of video generation models from [Kandinsky Lab](https://huggingface.co/kandinskylab). The
+main model generates video and synchronized audio from text or a reference image with a single multimodal diffusion
+transformer: video and audio latents are denoised together through fused blocks that cross-attend between the two
+modalities, each conditioned on its own Qwen2.5-VL text branch and a CLIP pooled embedding. A separate
+super-resolution model upscales the generated video tile by tile in the latent space of a causal 3D K-VAE.
+
+> [!TIP]
+> Check out the [Kandinsky Lab](https://huggingface.co/kandinskylab) organization on the Hub for the full set of
+> official checkpoints, including flow-matching and distilled variants of both the base and super-resolution models.
+>
+> Distilled checkpoints ship with the few-step [`PiflowScheduler`] and must be run with `guidance_scale=1.0`.
+
+## Available models
+
+| Model | Pipeline | Notes |
+|---|---|---|
+| [`kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers) | [`Kandinsky6TI2VAPipeline`] | Flow matching, `guidance_scale=5.0`, 50 steps |
+| [`kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers) | [`Kandinsky6TI2VAPipeline`] | Distilled, `guidance_scale=1.0`, 10 steps |
+| [`kandinskylab/Kandinsky-6.0-Lite-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-Lite-5s-Diffusers) | [`Kandinsky6TI2VAPipeline`] | Flow matching, `guidance_scale=5.0`, 50 steps |
+| [`kandinskylab/Kandinsky-6.0-Lite-distill-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-Lite-distill-5s-Diffusers) | [`Kandinsky6TI2VAPipeline`] | Distilled, `guidance_scale=1.0`, 10 steps |
+| [`kandinskylab/Kandinsky-6.0-Pro-pretrain-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-Pro-pretrain-5s-Diffusers) | [`Kandinsky6TI2VAPipeline`] | Flow matching, `guidance_scale=5.0`, 50 steps |
+| [`kandinskylab/Kandinsky-6.0-Lite-pretrain-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-Lite-pretrain-5s-Diffusers) | [`Kandinsky6TI2VAPipeline`] | Flow matching, `guidance_scale=5.0`, 50 steps |
+| [`kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-VSR-5s-Diffusers) | [`Kandinsky6SRPipeline`] | Flow matching super-resolution |
+| [`kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers`](https://huggingface.co/kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers) | [`Kandinsky6SRPipeline`] | Distilled super-resolution, 2 steps |
+
+## Text/image-to-video-and-audio
+
+```python
+import torch
+from diffusers import Kandinsky6TI2VAPipeline
+from diffusers.utils import encode_video
+
+pipe = Kandinsky6TI2VAPipeline.from_pretrained(
+    "kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers", torch_dtype=torch.bfloat16
+)
+pipe.enable_model_cpu_offload()
+
+output = pipe(
+    prompt="A cat and a dog baking a cake together in a kitchen.",
+    height=480,
+    width=864,
+    num_frames=121,
+    num_inference_steps=10,
+    guidance_scale=1.0,
+)
+encode_video(
+    output.frames[0],
+    fps=24,
+    output_path="output.mp4",
+    audio=output.audio[0][None],
+    audio_sample_rate=pipe.audio_sample_rate,
+)
+```
+
+Pass `image=` to condition the first frame on a reference image, `sample_audio=False` to generate video only, and
+`expand_prompts=True` to let the Qwen2.5-VL text encoder rewrite short prompts into detailed ones first.
+
+## Video super-resolution
+
+[`Kandinsky6SRPipeline`] takes the frames produced by [`Kandinsky6TI2VAPipeline`] and upscales them by `2`, `4`, or
+`2.25` (a 1.125x bilinear pre-upscale followed by the 2x path). The video is split into overlapping tiles, every tile
+is refined at one of the tile sizes the SR transformer was trained on, and the tiles are blended back with Hann
+windows.
+
+```python
+# required: lets inductor pick flex-attention tiles that fit the SR block mask
+torch._inductor.config.max_autotune = True  
+
+sr_pipe = Kandinsky6SRPipeline.from_pretrained(
+    "kandinskylab/Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers", torch_dtype=torch.bfloat16
+)
+# The SR transformer always runs NABLA sparse attention on the `flex` backend. Compile it, otherwise flex falls
+# back to an eager implementation that needs far more memory at video resolutions.
+sr_pipe.enable_model_cpu_offload()
+sr_pipe.transformer.set_attention_backend("flex")
+sr_pipe.transformer.compile_repeated_blocks(fullgraph=True)
+
+upscaled = sr_pipe(video=output.frames[0], resolution_scale=2.25, num_inference_steps=2).frames[0]
+```
+
+## Memory optimization
+
+Refer to the [Reduce memory usage](../../optimization/memory) guide for the general set of techniques. Both
+[`Kandinsky6TI2VAPipeline`] and [`Kandinsky6SRPipeline`] support [model offloading](../../optimization/memory#model-offloading)
+(used above) and, for a smaller footprint at the cost of speed, [sequential CPU offloading](../../optimization/memory#cpu-offloading):
+
+```python
+pipe.enable_sequential_cpu_offload()
+```
+
+[`Kandinsky6TI2VAPipeline`]'s video VAE also supports [tiled decoding](../../optimization/memory#vae-tiling) for high
+resolutions or long videos:
+
+```python
+pipe.vae.enable_tiling()
+```
+
+## Notes
+
+- `height` and `width` must be divisible by the video VAE's spatial compression ratio times the transformer's patch
+  size — `16` with the default [`Kandinsky6TI2VAPipeline`] configuration (`AutoencoderKLHunyuanVideo` at a
+  compression ratio of `8`, `patch_size=(1, 2, 2)`). `480x864`, used in the example above, satisfies this.
+- [`Kandinsky6SRPipeline`]'s input `video` must have `1 + k * vae_scale_factor_temporal` frames for some integer `k`
+  (a temporal compression ratio of `4` with the default K-VAE configuration, so `121` frames works but `120` doesn't)
+  — trim or pad a video that doesn't already satisfy this before upscaling it.
+- `expand_prompts=True` reuses the already-loaded Qwen2.5-VL text encoder for an extra generation pass before
+  denoising, so it adds latency but no extra model weights.
+- Compile the repeated transformer blocks for faster repeated inference:
+  ```python
+  pipe.transformer.compile_repeated_blocks(fullgraph=True)
+  ```
+
+## Kandinsky6TI2VAPipeline
+
+[[autodoc]] Kandinsky6TI2VAPipeline
+  - all
+  - __call__
+
+## Kandinsky6SRPipeline
+
+[[autodoc]] Kandinsky6SRPipeline
+  - all
+  - __call__
+
+## Kandinsky6SRLatentUpscalerBank
+
+[[autodoc]] Kandinsky6SRLatentUpscalerBank
+  - forward
+
+## Kandinsky6TI2VAPipelineOutput
+
+[[autodoc]] pipelines.kandinsky6.pipeline_output.Kandinsky6TI2VAPipelineOutput
+
+## Kandinsky6SRPipelineOutput
+
+[[autodoc]] pipelines.kandinsky6.pipeline_output.Kandinsky6SRPipelineOutput

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_xl.md

@@ -448,7 +448,7 @@ SDXL is a large model, and you may need to optimize memory to get it to run on y
 + refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)
 ```
 
-3. Enable [xFormers](../../../optimization/xformers) to run SDXL if `torch<2.0`:
+3. Enable [xFormers](../../../optimization/attention_backends) to run SDXL if `torch<2.0`:
 
 ```diff
 + base.enable_xformers_memory_efficient_attention()