Add Qwen-Image 2.1

docs/source/en/_toctree.yml

@@ -405,6 +405,8 @@
         title: PixArtTransformer2DModel
       - local: api/models/prior_transformer
         title: PriorTransformer
+      - local: api/models/qwenimage21_transformer2d
+        title: QwenImage21Transformer2DModel
       - local: api/models/qwenimage_transformer2d
         title: QwenImageTransformer2DModel
       - local: api/models/sana_transformer2d
@@ -489,6 +491,8 @@
         title: AutoencoderKLMochi
       - local: api/models/autoencoderkl_qwenimage
         title: AutoencoderKLQwenImage
+      - local: api/models/autoencoderkl_qwenimage21
+        title: AutoencoderKLQwenImage21
       - local: api/models/autoencoder_kl_wan
         title: AutoencoderKLWan
       - local: api/models/autoencoder_rae
@@ -637,6 +641,8 @@
         title: PRX
       - local: api/pipelines/prx_pixel
         title: PRX Pixel
+      - local: api/pipelines/qwenimage21
+        title: Qwen-Image 2.1
       - local: api/pipelines/qwenimage
         title: QwenImage
       - local: api/pipelines/sana

docs/source/en/api/models/autoencoderkl_qwenimage21.md

@@ -0,0 +1,37 @@
+<!-- Copyright 2026 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. -->
+
+# AutoencoderKLQwenImage21
+
+The 64-channel variational auto-encoder used by Qwen-Image 2.1. It compresses 16x spatially, and its per-channel
+`latents_mean` / `latents_std` are part of the config rather than a single scaling factor.
+
+```python
+import torch
+from diffusers import AutoencoderKLQwenImage21
+
+vae = AutoencoderKLQwenImage21.from_pretrained("Qwen/Qwen-Image-2.1", subfolder="vae", dtype=torch.bfloat16)
+```
+
+## AutoencoderKLQwenImage21
+
+[[autodoc]] AutoencoderKLQwenImage21
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/qwenimage21_transformer2d.md

@@ -0,0 +1,45 @@
+<!-- Copyright 2026 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. -->
+
+# QwenImage21Transformer2DModel
+
+The single-stream transformer used by Qwen-Image 2.1. Text and image latents share one sequence, and a single shared
+`modulation` projection feeds every block.
+
+Two behaviours distinguish 2.1 from earlier QwenImage transformers:
+
+- **Block-causal attention** — attention follows `(q_idx >= kv_idx) or same_image_block`, so the joint sequence is
+  causal while each image block stays internally bidirectional. `QwenImage21AttnProcessor` implements it as one
+  attention call per prefix segment and is the default. `QwenImage21FlexAttnProcessor` implements it as a single
+  `flex_attention` call driven by a `BlockMask`, which is faster once the model is compiled. Both produce the same
+  results.
+- `causal_condition` — text and condition-image tokens are modulated from `t = 0` rather than the sampled timestep.
+  Their activations are independent of the denoising step, so the keys and values of that prefix are cacheable
+  across steps via the `kv_cache` argument.
+
+Load it with:
+
+```python
+import torch
+from diffusers import QwenImage21Transformer2DModel
+
+transformer = QwenImage21Transformer2DModel.from_pretrained(
+    "Qwen/Qwen-Image-2.1", subfolder="transformer", dtype=torch.bfloat16
+)
+```
+
+## QwenImage21Transformer2DModel
+
+[[autodoc]] QwenImage21Transformer2DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/pipelines/qwenimage21.md

@@ -0,0 +1,73 @@
+<!-- Copyright 2026 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. -->
+
+# Qwen-Image 2.1
+
+Qwen-Image 2.1 encodes the prompt and any condition images together with a Qwen3-VL model, then denoises the target
+image with a single-stream block-causal transformer. See
+[`QwenImage21Transformer2DModel`](../models/qwenimage21_transformer2d) for block-causal attention, the attention
+processors, and `causal_condition`.
+
+The defaults are the values Qwen recommends: 40 steps and no guidance. Pass a `negative_prompt` together with
+`true_cfg_scale > 1` to turn classifier-free guidance on, which doubles the work per step.
+
+```python
+import torch
+from diffusers import QwenImage21Pipeline
+
+pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1", dtype=torch.bfloat16).to("cuda")
+
+# Text-to-image
+image = pipe("A capybara wearing a wizard hat, oil painting").images[0]
+image.save("t2i.png")
+
+# Image-conditioned editing
+edited = pipe("Move it to a snowy mountain top", image=image).images[0]
+edited.save("edit.png")
+```
+
+## Multiple condition images
+
+Pass a list to `image` and every entry becomes its own block in the joint sequence: the Qwen3-VL encoder sees them as
+vision context and the VAE contributes their latent tokens. Block-causal attention keeps each block internally
+bidirectional while letting later blocks and the target image attend to the earlier ones, so the order you pass them
+in is the order the model reads them.
+
+```python
+edited = pipe("Put the flowers from the first image into the second scene", image=[flowers, scene]).images[0]
+```
+
+## Faster attention with flex_attention
+
+The default `QwenImage21AttnProcessor` runs the block-causal prefill as one attention call per prefix segment. It
+needs no compilation and works on any PyTorch build. `QwenImage21FlexAttnProcessor` expresses the same mask as a
+single `flex_attention` call, which is faster once the model is **_compiled_**.
+
+> [!TIP]
+> Compile the model when you switch to the flex processor. An uncompiled `flex_attention` materializes the full
+> attention score matrix in fp32, which is much slower and runs out of memory at high resolution.
+
+```python
+from diffusers.models.transformers.transformer_qwenimage21 import QwenImage21FlexAttnProcessor
+
+pipe.transformer.set_attn_processor(QwenImage21FlexAttnProcessor())
+pipe.transformer.compile()
+```
+
+## QwenImage21Pipeline
+
+[[autodoc]] QwenImage21Pipeline
+    - all
+    - __call__
+
+## QwenImagePipelineOutput
+
+[[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

src/diffusers/__init__.py

@@ -253,6 +253,7 @@
             "AutoencoderKLMiniMaxH3Audio",
             "AutoencoderKLMochi",
             "AutoencoderKLQwenImage",
+            "AutoencoderKLQwenImage21",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderKLWan",
             "AutoencoderOobleck",
@@ -332,6 +333,7 @@
             "PixArtTransformer2DModel",
             "PriorTransformer",
             "PRXTransformer2DModel",
+            "QwenImage21Transformer2DModel",
             "QwenImageControlNetModel",
             "QwenImageMultiControlNetModel",
             "QwenImageTransformer2DModel",
@@ -764,6 +766,7 @@
             "PixArtSigmaPipeline",
             "PRXPipeline",
             "PRXPixelPipeline",
+            "QwenImage21Pipeline",
             "QwenImageControlNetInpaintPipeline",
             "QwenImageControlNetPipeline",
             "QwenImageEditInpaintPipeline",
@@ -1130,6 +1133,7 @@
             AutoencoderKLMiniMaxH3Audio,
             AutoencoderKLMochi,
             AutoencoderKLQwenImage,
+            AutoencoderKLQwenImage21,
             AutoencoderKLTemporalDecoder,
             AutoencoderKLWan,
             AutoencoderOobleck,
@@ -1209,6 +1213,7 @@
             PixArtTransformer2DModel,
             PriorTransformer,
             PRXTransformer2DModel,
+            QwenImage21Transformer2DModel,
             QwenImageControlNetModel,
             QwenImageMultiControlNetModel,
             QwenImageTransformer2DModel,
@@ -1616,6 +1621,7 @@
             PixArtSigmaPipeline,
             PRXPipeline,
             PRXPixelPipeline,
+            QwenImage21Pipeline,
             QwenImageControlNetInpaintPipeline,
             QwenImageControlNetPipeline,
             QwenImageEditInpaintPipeline,

src/diffusers/models/__init__.py

@@ -50,6 +50,7 @@
     _import_structure["autoencoders.autoencoder_kl_minimax_h3_audio"] = ["AutoencoderKLMiniMaxH3Audio"]
     _import_structure["autoencoders.autoencoder_kl_mochi"] = ["AutoencoderKLMochi"]
     _import_structure["autoencoders.autoencoder_kl_qwenimage"] = ["AutoencoderKLQwenImage"]
+    _import_structure["autoencoders.autoencoder_kl_qwenimage21"] = ["AutoencoderKLQwenImage21"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_kl_wan"] = ["AutoencoderKLWan"]
     _import_structure["autoencoders.autoencoder_longcat_audio_dit"] = ["LongCatAudioDiTVae"]
@@ -144,6 +145,7 @@
     _import_structure["transformers.transformer_ovis_image"] = ["OvisImageTransformer2DModel"]
     _import_structure["transformers.transformer_prx"] = ["PRXTransformer2DModel"]
     _import_structure["transformers.transformer_qwenimage"] = ["QwenImageTransformer2DModel"]
+    _import_structure["transformers.transformer_qwenimage21"] = ["QwenImage21Transformer2DModel"]
     _import_structure["transformers.transformer_sana_video"] = ["SanaVideoTransformer3DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
@@ -195,6 +197,7 @@
             AutoencoderKLMiniMaxH3Audio,
             AutoencoderKLMochi,
             AutoencoderKLQwenImage,
+            AutoencoderKLQwenImage21,
             AutoencoderKLTemporalDecoder,
             AutoencoderKLWan,
             AutoencoderOobleck,
@@ -287,6 +290,7 @@
             PixArtTransformer2DModel,
             PriorTransformer,
             PRXTransformer2DModel,
+            QwenImage21Transformer2DModel,
             QwenImageTransformer2DModel,
             SanaTransformer2DModel,
             SanaVideoTransformer3DModel,

src/diffusers/models/autoencoders/__init__.py

@@ -20,6 +20,7 @@
 from .autoencoder_kl_minimax_h3_audio import AutoencoderKLMiniMaxH3Audio
 from .autoencoder_kl_mochi import AutoencoderKLMochi
 from .autoencoder_kl_qwenimage import AutoencoderKLQwenImage
+from .autoencoder_kl_qwenimage21 import AutoencoderKLQwenImage21
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_kl_wan import AutoencoderKLWan
 from .autoencoder_longcat_audio_dit import LongCatAudioDiTVae