| ## Performance | |
| *Updated Feb 1st.* | |
| ### Low-level Question-Answering | |
| This model has reached 75.90\%(*13\% better than previous version*)/76.52\%(*10\% better than previous version*) on Q-Bench A1 *dev/test* (multi-choice questions). | |
| It also outperforms the following close-source models with much larger model capacities: | |
| | Model | *dev* | *test* | | |
| | ---- | ---- | ---- | | |
| | **Co-Instruct-Preview (mPLUG-Owl2) (This Model)** | **75.90\%** | **76.52\%** | | |
| | \*GPT-4V-Turbo | 74.41\% | 74.10\% | | |
| | \*Qwen-VL-**Max** | 73.63\% | 73.90\% | | |
| | \*GPT-4V (Nov. 2023) | 71.78\% | 73.44\% | | |
| | \*Gemini-Pro | 68.16\% | 69.46\% | | |
| | Q-Instruct (mPLUG-Owl2, Nov. 2023) | 67.42\% | 70.43\% | | |
| | \*Qwen-VL-Plus | 66.01\% | 68.93\% | | |
| | mPLUG-Owl2 | 62.14\% | 62.68\% | | |
| \*: Proprietary Models. | |
| #### Image/Video Quality Assessment | |
| | Model | live | agi | livec | test_spaq | csiq | test_kadid | test_koniq | konvid | maxwell_test | | |
| |--------------------------|--------------|--------------|-------------|-------------|-------------|-------------|-------------|-------------|--------------| | |
| |**Co-Instruct-Preview (mPLUG-Owl2) (This Model)** | **0.803/0.756** | **0.719**/0.732 | **0.827/0.835** | **0.946/0.937** | **0.711/0.727** | **0.782/0.766** | 0.886/**0.935** | **0.818/0.790** | **0.735/0.714** | | |
| | Q-Instruct (mPLUG-Owl2, Nov. 2023) | 0.749/0.747 | 0.710/**0.753** | 0.781/0.791 | 0.921/0.917 | 0.693/0.723 | 0.670/0.665 | **0.904**/0.921 | 0.766/0.738 | 0.650/0.649 | | |
| We are also constructing multi-image benchmark sets (image pairs, triple-quadruple images), and the results on multi-image benchmarks will be released soon! | |
| ## Load Model | |
| ```python | |
| import torch | |
| from transformers import AutoModelForCausalLM | |
| model = AutoModelForCausalLM.from_pretrained("q-future/co-instruct-preview", | |
| trust_remote_code=True, | |
| torch_dtype=torch.float16, | |
| attn_implementation="eager", | |
| device_map={"":"cuda:0"}) | |
| ``` | |
| ## Chat | |
| ```python | |
| import requests | |
| from PIL import Image | |
| ### Single Image | |
| prompt = "USER: The image: <|image|> Which happens in this image: motion-blur, over-exposure, or under-exposure? ASSISTANT:" | |
| url = "https://raw.githubusercontent.com/Q-Future/Q-Align/main/fig/singapore_flyer.jpg" | |
| image = Image.open(requests.get(url,stream=True).raw) | |
| model.chat(prompt, [image], max_new_tokens=200) | |
| ## Motion blur | |
| ### Double Image Comparison | |
| prompt_cmp = "USER: The first image: <|image|>\nThe second image: <|image|>Which image has better quality, and why? ASSISTANT:" | |
| url = "https://raw.githubusercontent.com/Q-Future/Q-Align/main/fig/boy_colorful.jpg" | |
| image_2 = Image.open(requests.get(url,stream=True).raw) | |
| model.chat(prompt_cmp, [image, image_2], max_new_tokens=200) | |
| ## The second image has better quality. The description indicates that the image has accurate exposure, precise focus, clear details, rich colors, and sufficient lighting. Additionally, the texture details are clear, and the composition is centered. In comparison, the first image has good clarity and rich texture details, but the lighting is slightly weak, which can affect the overall quality of the image. Therefore, the second image is of higher quality due to its accurate exposure, precise focus, clear details, rich colors, sufficient lighting, and centered composition. | |
| ``` |