---
license: apache-2.0
---

Modular Florence2 block that can also be used with Mellon.

# How to use

## With Mellon

The node can be used with the default installation of Mellon using the `Dynamic Block Node`

## Using it with code

### Captioning

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<CAPTION>"  # can also be <DETAILED_CAPTION> or <MORE_DETAILED_CAPTION>
annotation_prompt = ""

output = pipe(image=image, annotation_task=annotation_task, annotation_prompt=annotation_prompt).annotations[0]
print(output)
```

#### Caption

```
A man and a woman writing on a white board.
```

#### Detailed Caption

```
In this image we can see a man and a woman holding markers in their hands. We can also see a board with some text on it.
```

#### More Detailed Caption

```
A man and a woman are standing in front of a whiteboard. The woman is writing on a black marker. The man is wearing a blue shirt. The whiteboard has writing on it. The writing on the whiteboard is black. The people are looking at each other. There is writing in black marker on the board. There are drawings on whiteboard behind the people.
```

### Object Detection

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<OD>"
annotation_prompt = ""

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Output                                                                                                            |
| ------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------- |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Output](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/object_detection.png) |

### Dense Region Caption

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<DENSE_REGION_CAPTION>"
annotation_prompt = ""

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Output                                                                                                                |
| ------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------- |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Output](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/dense_region_caption.png) |

### Region Proposal

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<REGION_PROPOSAL>"
annotation_prompt = ""

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Output                                                                                                           |
| ------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------- |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Output](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/region_proposal.png) |

### Phrase Grounding

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<CAPTION_TO_PHRASE_GROUNDING>"
annotation_prompt = "man"

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",  # can also use `mask_image` and `mask_overlay`
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Bounding Box                                                                                                          | Mask Image                                                                                                            | Mask Overlay                                                                                                             |
| ------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/phrase_grounding_bbox.png) | ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/phrase_grounding_mask.png) | ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/phrase_grounding_overlay.png) |

### Referring Expression Segmentation

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<REFERRING_EXPRESSION_SEGMENTATION>"
annotation_prompt = "man"

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="mask_image",  # can also use `mask_overlay`
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Mask Image                                                                                                       | Mask Overlay                                                                                                        |
| ------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------- |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/ref_exp_seg_mask.png) | ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/ref_exp_seg_overlay.png) |

### Open Vocabulary Detection

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<OPEN_VOCABULARY_DETECTION>"
annotation_prompt = "man with a beard"

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Output                                                                                                           |
| ------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------- |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Output](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/open_vocabulary.png) |

### OCR

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<OCR>"
annotation_prompt = ""

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",
).annotations[0]
print(output)
```

```
The Diffuser's library byHugging Face makes it easyfor developers to run imagegeneration and influenceusing state-of-the-astdiffusion models withjust a few lines of codehuman eou
```

### OCR with region

```python
import torch

from diffusers.modular_pipelines import ModularPipeline
from diffusers.utils import load_image


pipe = ModularPipeline.from_pretrained("OzzyGT/florence-2-block", trust_remote_code=True)
pipe.load_components(torch_dtype=torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png"
)

annotation_task = "<OCR_WITH_REGION>"
annotation_prompt = ""

output = pipe(
    image=image,
    annotation_task=annotation_task,
    annotation_prompt=annotation_prompt,
    annotation_output_type="bounding_box",
).images[0]
output.save("output.png")
```

| Input                                                                                                              | Output                                                                                                      |
| ------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------- |
| ![Input](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/white_board_people.png) | ![Output](https://huggingface.co/datasets/OzzyGT/diffusers-examples/resolve/main/florence-2/ocr_region.png) |