Upload README.md

README.md

@@ -52,7 +52,7 @@ blocks = ModularPipelineBlocks.from_pretrained(
     trust_components_code=True,
 )
 pipe = blocks.init_pipeline()
-pipe.load_components(dtype={"default": torch.bfloat16, "vae": torch.float32})
+pipe.load_components(workflow="text2image", dtype={"default": torch.bfloat16, "vae": torch.float32})
 pipe.to("cuda")
 
 layout = {
@@ -124,6 +124,44 @@ image = pipe(
 image.save("moss_image.png")
 ```
 
+## Reference images
+
+MOSS-Image can also take reference images, through the vision part of its Qwen3.5 text encoder. The model wasn't trained for this: references guide the content, style and layout of the image, not exact pixels or identity. Load the `reference` workflow and pass the images with `reference_images`:
+
+```python
+import torch
+from diffusers import ModularPipelineBlocks
+from diffusers.utils import load_image
+
+
+blocks = ModularPipelineBlocks.from_pretrained(
+    "OzzyGT/moss_image_custom_blocks",
+    trust_remote_code=True,
+    components_repo="OzzyGT/MOSS_Image_SFT_diffusers",
+    trust_components_code=True,
+)
+pipe = blocks.init_pipeline()
+pipe.load_components(workflow="reference", dtype={"default": torch.bfloat16, "vae": torch.float32})
+pipe.to("cuda")
+
+reference = load_image(
+    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/krea2_reference/20260907154202_edit_image_0.png"
+)
+
+image = pipe(
+    prompt="photo",
+    reference_images=reference,
+    num_inference_steps=28,
+    height=1024,
+    width=1024,
+    generator=torch.Generator("cuda").manual_seed(42),
+    output="images",
+)[0]
+image.save("moss_image_reference.png")
+```
+
+`reference_max_pixels` (default 1024 x 1024) sets how large each reference is for the vision encoder. To load only what text-to-image needs, without the vision encoder, use `workflow="text2image"` as in the sample above.
+
 ## License
 
 This model is released under the [Apache-2.0 license](LICENSE), the same as the original.