update

Author: DN6

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

@@ -21,6 +21,29 @@ from diffusers import HiDreamImageTransformer2DModel
 transformer = HiDreamImageTransformer2DModel.from_pretrained("HiDream-ai/HiDream-I1-Full", subfolder="transformer", torch_dtype=torch.bfloat16)
 ```
 
+## Loading GGUF quantized checkpoints
+
+GGUF checkpoints for the `HiDreamImageTransformer2DModel` can we be loaded using `~FromOriginalModelMixin.from_single_file`
+
+```python
+from diffusers import HiDreamImageTransformer2DModel
+
+ckpt_path = "https://huggingface.co/city96/HiDream-I1-Dev-gguf/blob/main/hidream-i1-dev-Q2_K.gguf"
+transformer = HiDreamImageTransformer2DModel.from_single_file(ckpt_path, torch_dtype=torch.bfloat16)
+```
+
+If you are trying to use a GGUF checkpoint for the `HiDream-ai/HiDream-E1-Full` model, you will have to pass in a `config` argument to properly configure the model. This is because the HiDream I1 and E1 models share the same state dict keys, so it is currently not possible to automatically infer the model type from the checkpoint itself.
+
+```python
+from diffusers import HiDreamImageTransformer2DModel
+
+ckpt_path = "https://huggingface.co/ND911/HiDream_e1_full_bf16-ggufs/blob/main/hidream_e1_full_bf16-Q2_K.gguf"
+
+transformer = HiDreamImageTransformer2DModel.from_single_file(ckpt_path, config="HiDream-ai/HiDream-E1-Full", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+
+
 ## HiDreamImageTransformer2DModel
 
 [[autodoc]] HiDreamImageTransformer2DModel

src/diffusers/loaders/single_file_model.py

@@ -31,6 +31,7 @@
     convert_autoencoder_dc_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
     convert_flux_transformer_checkpoint_to_diffusers,
+    convert_hidream_transformer_to_diffusers,
     convert_hunyuan_video_transformer_to_diffusers,
     convert_ldm_unet_checkpoint,
     convert_ldm_vae_checkpoint,
@@ -133,6 +134,10 @@
         "checkpoint_mapping_fn": convert_wan_vae_to_diffusers,
         "default_subfolder": "vae",
     },
+    "HiDreamImageTransformer2DModel": {
+        "checkpoint_mapping_fn": convert_hidream_transformer_to_diffusers,
+        "default_subfolder": "transformer",
+    },
 }
 
 

src/diffusers/loaders/single_file_utils.py

@@ -126,6 +126,7 @@
     ],
     "wan": ["model.diffusion_model.head.modulation", "head.modulation"],
     "wan_vae": "decoder.middle.0.residual.0.gamma",
+    "hidream": "double_stream_blocks.0.block.adaLN_modulation.1.bias",
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -190,6 +191,7 @@
     "wan-t2v-1.3B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"},
     "wan-t2v-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-14B-Diffusers"},
     "wan-i2v-14B": {"pretrained_model_name_or_path": "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"},
+    "hidream": {"pretrained_model_name_or_path": "HiDream-ai/HiDream-I1-Dev"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -701,6 +703,8 @@ def infer_diffusers_model_type(checkpoint):
     elif CHECKPOINT_KEY_NAMES["wan_vae"] in checkpoint:
         # All Wan models use the same VAE so we can use the same default model repo to fetch the config
         model_type = "wan-t2v-14B"
+    elif CHECKPOINT_KEY_NAMES["hidream"] in checkpoint:
+        model_type = "hidream"
     else:
         model_type = "v1"
 
@@ -3293,3 +3297,12 @@ def convert_wan_vae_to_diffusers(checkpoint, **kwargs):
             converted_state_dict[key] = value
 
     return converted_state_dict
+
+
+def convert_hidream_transformer_to_diffusers(checkpoint, **kwargs):
+    keys = list(checkpoint.keys())
+    for k in keys:
+        if "model.diffusion_model." in k:
+            checkpoint[k.replace("model.diffusion_model.", "")] = checkpoint.pop(k)
+
+    return checkpoint