Merge branch 'main' of https://huggingface.co/spaces/OpenGenAI/parti-prompts-leaderboard

app.py

@@ -23,8 +23,10 @@ MODEL_KEYS = "-".join(SUBMISSIONS.keys())
 SUBMISSION_ORG = f"results-{MODEL_KEYS}"
 
 submission_names = list(SUBMISSIONS.keys())
-parti_prompt_categories = load_dataset(os.path.join(ORG, "sd-v1-5"))["train"]["Category"]
-parti_prompt_challenge = load_dataset(os.path.join(ORG, "sd-v1-5"))["train"]["Challenge"]
+ds = load_dataset("nateraw/parti-prompts")["train"]
+
+parti_prompt_categories = ds["Category"]
+parti_prompt_challenge = ds["Challenge"]
 
 
 def load_submissions():
@@ -86,15 +88,26 @@ def get_dataframe_all():
 
 TITLE = "# Open Parti Prompts Leaderboard"
 DESCRIPTION = """
-*This leaderboard is retrieved from answers of [Community Evaluations on Parti Prompts](https://huggingface.co/spaces/OpenGenAI/open-parti-prompts)*
+The *Open Parti Prompts Leaderboard* compares state-of-the-art, open-source text-to-image models to each other according to **human preferences**. \n\n
+Text-to-image models are notoriously difficult to evaluate. [FID](https://en.wikipedia.org/wiki/Fr%C3%A9chet_inception_distance) and 
+[CLIP Score](https://en.wikipedia.org/wiki/Fr%C3%A9chet_inception_distance) are not enough to accurately state whether a text-to-image model can 
+**generate "good" images**. "Good" is extremely difficult to put into numbers. \n\n
+Instead, the **Open Parti Prompts Leaderboard** uses human feedback from the community to compare images from different text-to-image models to each other.
+
+\n\n
+
+❤️ ***Please take 3 minutes to contribute to the benchmark.*** \n
+👉 ***Play one round of [Open Parti Prompts Game](https://huggingface.co/spaces/OpenGenAI/open-parti-prompts) to contribute 10 answers.*** 🤗
 """
 
 EXPLANATION = """\n\n
 ## How the is data collected 📊 \n\n
 
-In the [Community Parti Prompts](https://huggingface.co/spaces/OpenGenAI/open-parti-prompts), community members select for every prompt
-of [Parti Prompts](https://huggingface.co/datasets/nateraw/parti-prompts) which open-source image generation model has generated the best image.
-The community's answers are then stored and used in this space to give a human evaluation of the different models. \n\n
+In more detail, the [Open Parti Prompts Game](https://huggingface.co/spaces/OpenGenAI/open-parti-prompts) collects human preferences that state which generated image 
+best fits a given prompt from the [Parti Prompts](https://huggingface.co/datasets/nateraw/parti-prompts) dataset. Parti Prompts has been designed to challenge
+text-to-image models on prompts of varying categories and difficulty. The images have been pre-generated from the models that are compared in this space.
+For more information of how the images were created, please refer to [Open Parti Prompts](https://huggingface.co/spaces/OpenGenAI/open-parti-prompts).
+The community's answers are then stored and used in this space to give a human-preference-based comparison of the different models. \n\n
 
 Currently the leaderboard includes the following models:
 - [sd-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5)
@@ -102,7 +115,8 @@ Currently the leaderboard includes the following models:
 - [if-v1-0](https://huggingface.co/DeepFloyd/IF-I-XL-v1.0)
 - [karlo](https://huggingface.co/kakaobrain/karlo-v1-alpha) \n\n
 
-In the following you can see three result tables. The first shows you the overall preferences across all prompts. The second and third tables
+In the following you can see three result tables. The first shows the overall comparison of the 4 models. The score states, 
+**the percentage at which images generated from the corresponding model are preferred over the image from all other models**. The second and third tables
 show you a breakdown analysis per category and per type of challenge as defined by [Parti Prompts](https://huggingface.co/datasets/nateraw/parti-prompts).
 """