README.md
| 1 | --- |
| 2 | license: other |
| 3 | license_name: bria-rmbg-1.4 |
| 4 | license_link: https://bria.ai/bria-huggingface-model-license-agreement/ |
| 5 | pipeline_tag: image-segmentation |
| 6 | tags: |
| 7 | - remove background |
| 8 | - background |
| 9 | - background-removal |
| 10 | - Pytorch |
| 11 | - vision |
| 12 | - legal liability |
| 13 | - transformers |
| 14 | - transformers.js |
| 15 | |
| 16 | extra_gated_description: RMBG v1.4 is available as a source-available model for non-commercial use |
| 17 | extra_gated_heading: "Fill in this form to get instant access" |
| 18 | extra_gated_fields: |
| 19 | Name: text |
| 20 | Company/Org name: text |
| 21 | Org Type (Early/Growth Startup, Enterprise, Academy): text |
| 22 | Role: text |
| 23 | Country: text |
| 24 | Email: text |
| 25 | By submitting this form, I agree to BRIA’s Privacy policy and Terms & conditions, see links below: checkbox |
| 26 | --- |
| 27 | |
| 28 | # BRIA Background Removal v1.4 Model Card |
| 29 | |
| 30 | RMBG v1.4 is our state-of-the-art background removal model, designed to effectively separate foreground from background in a range of |
| 31 | categories and image types. This model has been trained on a carefully selected dataset, which includes: |
| 32 | general stock images, e-commerce, gaming, and advertising content, making it suitable for commercial use cases powering enterprise content creation at scale. |
| 33 | The accuracy, efficiency, and versatility currently rival leading source-available models. |
| 34 | It is ideal where content safety, legally licensed datasets, and bias mitigation are paramount. |
| 35 | |
| 36 | Developed by BRIA AI, RMBG v1.4 is available as a source-available model for non-commercial use. |
| 37 | |
| 38 | |
| 39 | To purchase a commercial license, simply click [Here](https://go.bria.ai/3D5EGp0). |
| 40 | |
| 41 | |
| 42 | [CLICK HERE FOR A DEMO](https://huggingface.co/spaces/briaai/BRIA-RMBG-1.4) |
| 43 | |
| 44 | **NOTE** New RMBG version available! Check out [RMBG-2.0](https://huggingface.co/briaai/RMBG-2.0) |
| 45 | |
| 46 | Join our [Discord community](https://discord.gg/Nxe9YW9zHS) for more information, tutorials, tools, and to connect with other users! |
| 47 | |
| 48 | |
| 49 |  |
| 50 | |
| 51 | |
| 52 | ### Model Description |
| 53 | |
| 54 | - **Developed by:** [BRIA AI](https://bria.ai/) |
| 55 | - **Model type:** Background Removal |
| 56 | - **License:** [bria-rmbg-1.4](https://bria.ai/bria-huggingface-model-license-agreement/) |
| 57 | - The model is released under a Creative Commons license for non-commercial use. |
| 58 | - Commercial use is subject to a commercial agreement with BRIA. To purchase a commercial license simply click [Here](https://go.bria.ai/3B4Asxv). |
| 59 | |
| 60 | - **Model Description:** BRIA RMBG 1.4 is a saliency segmentation model trained exclusively on a professional-grade dataset. |
| 61 | - **BRIA:** Resources for more information: [BRIA AI](https://bria.ai/) |
| 62 | |
| 63 | |
| 64 | |
| 65 | ## Training data |
| 66 | Bria-RMBG model was trained with over 12,000 high-quality, high-resolution, manually labeled (pixel-wise accuracy), fully licensed images. |
| 67 | Our benchmark included balanced gender, balanced ethnicity, and people with different types of disabilities. |
| 68 | For clarity, we provide our data distribution according to different categories, demonstrating our model’s versatility. |
| 69 | |
| 70 | ### Distribution of images: |
| 71 | |
| 72 | | Category | Distribution | |
| 73 | | -----------------------------------| -----------------------------------:| |
| 74 | | Objects only | 45.11% | |
| 75 | | People with objects/animals | 25.24% | |
| 76 | | People only | 17.35% | |
| 77 | | people/objects/animals with text | 8.52% | |
| 78 | | Text only | 2.52% | |
| 79 | | Animals only | 1.89% | |
| 80 | |
| 81 | | Category | Distribution | |
| 82 | | -----------------------------------| -----------------------------------------:| |
| 83 | | Photorealistic | 87.70% | |
| 84 | | Non-Photorealistic | 12.30% | |
| 85 | |
| 86 | |
| 87 | | Category | Distribution | |
| 88 | | -----------------------------------| -----------------------------------:| |
| 89 | | Non Solid Background | 52.05% | |
| 90 | | Solid Background | 47.95% |
| 91 | |
| 92 | |
| 93 | | Category | Distribution | |
| 94 | | -----------------------------------| -----------------------------------:| |
| 95 | | Single main foreground object | 51.42% | |
| 96 | | Multiple objects in the foreground | 48.58% | |
| 97 | |
| 98 | |
| 99 | ## Qualitative Evaluation |
| 100 | |
| 101 |  |
| 102 | |
| 103 | |
| 104 | ## Architecture |
| 105 | |
| 106 | RMBG v1.4 is developed on the [IS-Net](https://github.com/xuebinqin/DIS) enhanced with our unique training scheme and proprietary dataset. |
| 107 | These modifications significantly improve the model’s accuracy and effectiveness in diverse image-processing scenarios. |
| 108 | |
| 109 | ## Installation |
| 110 | ```bash |
| 111 | pip install -qr https://huggingface.co/briaai/RMBG-1.4/resolve/main/requirements.txt |
| 112 | ``` |
| 113 | |
| 114 | ## Usage |
| 115 | |
| 116 | Either load the pipeline |
| 117 | ```python |
| 118 | from transformers import pipeline |
| 119 | image_path = "https://farm5.staticflickr.com/4007/4322154488_997e69e4cf_z.jpg" |
| 120 | pipe = pipeline("image-segmentation", model="briaai/RMBG-1.4", trust_remote_code=True) |
| 121 | pillow_mask = pipe(image_path, return_mask = True) # outputs a pillow mask |
| 122 | pillow_image = pipe(image_path) # applies mask on input and returns a pillow image |
| 123 | ``` |
| 124 | |
| 125 | Or load the model |
| 126 | ```python |
| 127 | from PIL import Image |
| 128 | from skimage import io |
| 129 | import torch |
| 130 | import torch.nn.functional as F |
| 131 | from transformers import AutoModelForImageSegmentation |
| 132 | from torchvision.transforms.functional import normalize |
| 133 | model = AutoModelForImageSegmentation.from_pretrained("briaai/RMBG-1.4",trust_remote_code=True) |
| 134 | def preprocess_image(im: np.ndarray, model_input_size: list) -> torch.Tensor: |
| 135 | if len(im.shape) < 3: |
| 136 | im = im[:, :, np.newaxis] |
| 137 | # orig_im_size=im.shape[0:2] |
| 138 | im_tensor = torch.tensor(im, dtype=torch.float32).permute(2,0,1) |
| 139 | im_tensor = F.interpolate(torch.unsqueeze(im_tensor,0), size=model_input_size, mode='bilinear') |
| 140 | image = torch.divide(im_tensor,255.0) |
| 141 | image = normalize(image,[0.5,0.5,0.5],[1.0,1.0,1.0]) |
| 142 | return image |
| 143 | |
| 144 | def postprocess_image(result: torch.Tensor, im_size: list)-> np.ndarray: |
| 145 | result = torch.squeeze(F.interpolate(result, size=im_size, mode='bilinear') ,0) |
| 146 | ma = torch.max(result) |
| 147 | mi = torch.min(result) |
| 148 | result = (result-mi)/(ma-mi) |
| 149 | im_array = (result*255).permute(1,2,0).cpu().data.numpy().astype(np.uint8) |
| 150 | im_array = np.squeeze(im_array) |
| 151 | return im_array |
| 152 | |
| 153 | device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") |
| 154 | model.to(device) |
| 155 | |
| 156 | # prepare input |
| 157 | image_path = "https://farm5.staticflickr.com/4007/4322154488_997e69e4cf_z.jpg" |
| 158 | orig_im = io.imread(image_path) |
| 159 | orig_im_size = orig_im.shape[0:2] |
| 160 | model_input_size = [1024, 1024] |
| 161 | image = preprocess_image(orig_im, model_input_size).to(device) |
| 162 | |
| 163 | # inference |
| 164 | result=model(image) |
| 165 | |
| 166 | # post process |
| 167 | result_image = postprocess_image(result[0][0], orig_im_size) |
| 168 | |
| 169 | # save result |
| 170 | pil_mask_im = Image.fromarray(result_image) |
| 171 | orig_image = Image.open(image_path) |
| 172 | no_bg_image = orig_image.copy() |
| 173 | no_bg_image.putalpha(pil_mask_im) |
| 174 | ``` |
| 175 | |
| 176 | |