Add MAGI-1: Autoregressive Video Generation at Scale

docs/source/en/_toctree.yml

@@ -315,6 +315,8 @@
         title: Lumina2Transformer2DModel
       - local: api/models/lumina_nextdit2d
         title: LuminaNextDiT2DModel
+      - local: api/models/magi_transformer_3d
+        title: MagiTransformer3DModel
       - local: api/models/mochi_transformer3d
         title: MochiTransformer3DModel
       - local: api/models/omnigen_transformer
@@ -369,6 +371,8 @@
         title: AutoencoderKLHunyuanVideo
       - local: api/models/autoencoderkl_ltx_video
         title: AutoencoderKLLTXVideo
+      - local: api/models/autoencoder_kl_magi
+        title: AutoencoderKLMagi
       - local: api/models/autoencoderkl_magvit
         title: AutoencoderKLMagvit
       - local: api/models/autoencoderkl_mochi
@@ -491,6 +495,8 @@
       title: Lumina 2.0
     - local: api/pipelines/lumina
       title: Lumina-T2X
+    - local: api/pipelines/magi
+      title: MAGI-1
     - local: api/pipelines/marigold
       title: Marigold
     - local: api/pipelines/mochi

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

@@ -0,0 +1,34 @@
+<!-- Copyright 2024 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. -->
+
+# AutoencoderKLMagi
+
+The 3D variational autoencoder (VAE) model with KL loss used in [MAGI-1: Autoregressive Video Generation at Scale](https://arxiv.org/abs/2505.13211) by Sand.ai.
+
+MAGI-1 uses a transformer-based VAE with 8x spatial and 4x temporal compression, providing fast average decoding time and highly competitive reconstruction quality.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLMagi
+
+vae = AutoencoderKLMagi.from_pretrained("sand-ai/MAGI-1", subfolder="vae", torch_dtype=torch.float32)
+```
+
+## AutoencoderKLMagi
+
+[[autodoc]] AutoencoderKLMagi
+  - decode
+  - all
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/magi_transformer_3d.md

@@ -0,0 +1,32 @@
+<!-- Copyright 2024 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. -->
+
+# MagiTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [MAGI-1: Autoregressive Video Generation at Scale](https://arxiv.org/abs/2505.13211) by Sand.ai.
+
+MAGI-1 is an autoregressive denoising video generation model that generates videos chunk-by-chunk instead of as a whole. Each chunk (24 frames) is denoised holistically, and the generation of the next chunk begins as soon as the current one reaches a certain level of denoising.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import MagiTransformer3DModel
+
+transformer = MagiTransformer3DModel.from_pretrained("sand-ai/MAGI-1", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## MagiTransformer3DModel
+
+[[autodoc]] MagiTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput