Add Hy-Nomas support with a modular pipeline

docs/source/en/_toctree.yml

@@ -373,6 +373,8 @@
         title: HunyuanVideo15Transformer3DModel
       - local: api/models/hunyuan_video_transformer_3d
         title: HunyuanVideoTransformer3DModel
+      - local: api/models/hy_nomas_transformer3d
+        title: HyNomasTransformer3DModel
       - local: api/models/ideogram4_transformer2d
         title: Ideogram4Transformer2DModel
       - local: api/models/transformer_joyimage_edit_plus
@@ -523,6 +525,8 @@
         title: LTX2VideoDiffusionDecoderModel
       - local: api/models/autoencoder_oobleck
         title: Oobleck AutoEncoder
+      - local: api/models/pixelumm_decoder
+        title: PixelUMMDecoder
       - local: api/models/autoencoder_tiny
         title: Tiny AutoEncoder
       - local: api/models/vq
@@ -751,6 +755,8 @@
         title: HunyuanVideo
       - local: api/pipelines/hunyuan_video15
         title: HunyuanVideo1.5
+      - local: api/pipelines/hy_nomas
+        title: Hy-Nomas
       - local: api/pipelines/kandinsky5_video
         title: Kandinsky 5.0 Video
       - local: api/pipelines/kandinsky6

docs/source/en/api/models/hy_nomas_transformer3d.md

@@ -0,0 +1,30 @@
+<!-- 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. -->
+
+# HyNomasTransformer3DModel
+
+A double-stream transformer for segment autoregressive video generation. Video tokens use three-dimensional rotary
+positions, text tokens retain their segment-local positions, and camera attention conditions each latent frame on
+four pixel-frame extrinsics. The caller supplies the key/value cache explicitly.
+
+```python
+import torch
+from diffusers import HyNomasTransformer3DModel
+
+model = HyNomasTransformer3DModel.from_pretrained(
+    "tencent/Hy-Nomas-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16
+)
+```
+
+## HyNomasTransformer3DModel
+
+[[autodoc]] HyNomasTransformer3DModel
+    - forward

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

@@ -0,0 +1,33 @@
+<!-- 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. -->
+
+# PixelUMMDecoder
+
+PixelUMM is a distilled diffusion decoder for 32-channel video latents, with spatial compression 32 and temporal
+compression 4. It predicts clean pixels conditioned on the latent video and an independently sampled noise tensor.
+It does not encode videos into latents.
+
+```python
+import torch
+from diffusers import PixelUMMDecoder
+
+decoder = PixelUMMDecoder.from_pretrained("tencent/Hy-Nomas-Diffusers", subfolder="decoder", torch_dtype=torch.bfloat16)
+```
+
+The decoder supports local spatial attention, overlapping temporal windows, optional spatial tiling, and explicit
+streaming sessions. Noise is indexed by global latent frame, allowing overlapping windows to use consistent noise.
+
+## PixelUMMDecoder
+
+[[autodoc]] PixelUMMDecoder
+    - forward
+    - decode
+    - create_streaming_decode_session