README.md
| 1 | # Model Card for PickScore v1 |
| 2 | |
| 3 | This model is a scoring function for images generated from text. It takes as input a prompt and a generated image and outputs a score. |
| 4 | It can be used as a general scoring function, and for tasks such as human preference prediction, model evaluation, image ranking, and more. |
| 5 | See our paper [Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation](https://arxiv.org/abs/2305.01569) for more details. |
| 6 | |
| 7 | |
| 8 | ## Model Details |
| 9 | |
| 10 | ### Model Description |
| 11 | |
| 12 | This model was finetuned from CLIP-H using the [Pick-a-Pic dataset](https://huggingface.co/datasets/yuvalkirstain/pickapic_v1). |
| 13 | |
| 14 | ### Model Sources [optional] |
| 15 | |
| 16 | <!-- Provide the basic links for the model. --> |
| 17 | |
| 18 | - **Repository:** [See the PickScore repo](https://github.com/yuvalkirstain/PickScore) |
| 19 | - **Paper:** [Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation](https://arxiv.org/abs/2305.01569). |
| 20 | - **Demo [optional]:** [Huggingface Spaces demo for PickScore](https://huggingface.co/spaces/yuvalkirstain/PickScore) |
| 21 | |
| 22 | ## How to Get Started with the Model |
| 23 | |
| 24 | Use the code below to get started with the model. |
| 25 | |
| 26 | ```python |
| 27 | # import |
| 28 | from transformers import AutoProcessor, AutoModel |
| 29 | |
| 30 | # load model |
| 31 | device = "cuda" |
| 32 | processor_name_or_path = "laion/CLIP-ViT-H-14-laion2B-s32B-b79K" |
| 33 | model_pretrained_name_or_path = "yuvalkirstain/PickScore_v1" |
| 34 | |
| 35 | processor = AutoProcessor.from_pretrained(processor_name_or_path) |
| 36 | model = AutoModel.from_pretrained(model_pretrained_name_or_path).eval().to(device) |
| 37 | |
| 38 | def calc_probs(prompt, images): |
| 39 | |
| 40 | # preprocess |
| 41 | image_inputs = processor( |
| 42 | images=images, |
| 43 | padding=True, |
| 44 | truncation=True, |
| 45 | max_length=77, |
| 46 | return_tensors="pt", |
| 47 | ).to(device) |
| 48 | |
| 49 | text_inputs = processor( |
| 50 | text=prompt, |
| 51 | padding=True, |
| 52 | truncation=True, |
| 53 | max_length=77, |
| 54 | return_tensors="pt", |
| 55 | ).to(device) |
| 56 | |
| 57 | |
| 58 | with torch.no_grad(): |
| 59 | # embed |
| 60 | image_embs = model.get_image_features(**image_inputs) |
| 61 | image_embs = image_embs / torch.norm(image_embs, dim=-1, keepdim=True) |
| 62 | |
| 63 | text_embs = model.get_text_features(**text_inputs) |
| 64 | text_embs = text_embs / torch.norm(text_embs, dim=-1, keepdim=True) |
| 65 | |
| 66 | # score |
| 67 | scores = model.logit_scale.exp() * (text_embs @ image_embs.T)[0] |
| 68 | |
| 69 | # get probabilities if you have multiple images to choose from |
| 70 | probs = torch.softmax(scores, dim=-1) |
| 71 | |
| 72 | return probs.cpu().tolist() |
| 73 | |
| 74 | pil_images = [Image.open("my_amazing_images/1.jpg"), Image.open("my_amazing_images/2.jpg")] |
| 75 | prompt = "fantastic, increadible prompt" |
| 76 | print(calc_probs(prompt, pil_images)) |
| 77 | ``` |
| 78 | ## Training Details |
| 79 | |
| 80 | ### Training Data |
| 81 | |
| 82 | This model was trained on the [Pick-a-Pic dataset](https://huggingface.co/datasets/yuvalkirstain/pickapic_v1). |
| 83 | |
| 84 | |
| 85 | ### Training Procedure |
| 86 | |
| 87 | TODO - add paper. |
| 88 | |
| 89 | |
| 90 | ## Citation [optional] |
| 91 | |
| 92 | If you find this work useful, please cite: |
| 93 | |
| 94 | ```bibtex |
| 95 | @inproceedings{Kirstain2023PickaPicAO, |
| 96 | title={Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation}, |
| 97 | author={Yuval Kirstain and Adam Polyak and Uriel Singer and Shahbuland Matiana and Joe Penna and Omer Levy}, |
| 98 | year={2023} |
| 99 | } |
| 100 | ``` |
| 101 | |
| 102 | **APA:** |
| 103 | |
| 104 | [More Information Needed] |
| 105 | |
| 106 | |
| 107 | |