Almost-AGI-Diffusion/kand2

---
dataset_info:
features:
- name: Prompt
dtype: string
- name: Category
dtype: string
- name: Challenge
dtype: string
- name: Note
dtype: string
- name: images
dtype: image
- name: model_name
dtype: string
- name: seed
dtype: int64
- name: upvotes
dtype: int64
splits:
- name: train
num_bytes: 21708501.0
num_examples: 219
download_size: 21693707
dataset_size: 21708501.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---

# Kandinksy 2.2

All images included in this dataset were voted as "Not solved" by the community in https://huggingface.co/spaces/OpenGenAI/open-parti-prompts. This means that according to the community the model did not generate an image that corresponds sufficiently enough to the prompt.

The following script was used to generate the images:

```py
import PIL
import torch
from datasets import Dataset, Features
from datasets import Image as ImageFeature
from datasets import Value, load_dataset

from diffusers import DiffusionPipeline

def main():
print("Loading dataset...")
parti_prompts = load_dataset("nateraw/parti-prompts", split="train")

print("Loading pipeline...")
pipe_prior = DiffusionPipeline.from_pretrained(
"kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
)
pipe_prior.to("cuda")
pipe_prior.set_progress_bar_config(disable=True)

t2i_pipe = DiffusionPipeline.from_pretrained(
"kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16
)
t2i_pipe.to("cuda")
t2i_pipe.set_progress_bar_config(disable=True)

seed = 0
generator = torch.Generator("cuda").manual_seed(seed)
ckpt_id = (
"kandinsky-community/" + "kandinsky-2-2-prior" + "_" + "kandinsky-2-2-decoder"
)

print("Running inference...")
main_dict = {}
for i in range(len(parti_prompts)):
sample = parti_prompts[i]
prompt = sample["Prompt"]

image_embeds, negative_image_embeds = pipe_prior(
prompt,
generator=generator,
num_inference_steps=100,
guidance_scale=7.5,
).to_tuple()
image = t2i_pipe(
image_embeds=image_embeds,
negative_image_embeds=negative_image_embeds,
generator=generator,
num_inference_steps=100,
guidance_scale=7.5,
).images[0]

image = image.resize((256, 256), resample=PIL.Image.Resampling.LANCZOS)
img_path = f"kandinsky_22_{i}.png"
image.save(img_path)
main_dict.update(
{
prompt: {
"img_path": img_path,
"Category": sample["Category"],
"Challenge": sample["Challenge"],
"Note": sample["Note"],
"model_name": ckpt_id,
"seed": seed,
}
}
)

def generation_fn():
for prompt in main_dict:
prompt_entry = main_dict[prompt]
yield {
"Prompt": prompt,
"Category": prompt_entry["Category"],
"Challenge": prompt_entry["Challenge"],
"Note": prompt_entry["Note"],
"images": {"path": prompt_entry["img_path"]},
"model_name": prompt_entry["model_name"],
"seed": prompt_entry["seed"],
}

print("Preparing HF dataset...")
ds = Dataset.from_generator(
generation_fn,
features=Features(
Prompt=Value("string"),
Category=Value("string"),
Challenge=Value("string"),
Note=Value("string"),
images=ImageFeature(),
model_name=Value("string"),
seed=Value("int64"),
),
)
ds_id = "diffusers-parti-prompts/kandinsky-2-2"
ds.push_to_hub(ds_id)

if __name__ == "__main__":
main()
```

This dataset contains images generated by the Kandinsky 2.2 model, which were voted as Not solved by the community, indicating that the generated images do not sufficiently correspond to the given prompts. The dataset features include Prompt, Category, Challenge, Note, images, model_name, seed, and upvotes. The dataset was generated using a specific Python script that employs the Kandinsky 2.2 model to create images and uploads the results to the HuggingFace Hub.

原始信息汇总

数据集概述

数据集信息

• 特征列表:

  • Prompt: 字符串类型
  • Category: 字符串类型
  • Challenge: 字符串类型
  • Note: 字符串类型
  • images: 图像类型
  • model_name: 字符串类型
  • seed: 整数类型 (int64)
  • upvotes: 整数类型 (int64)

• 数据分割:

  • train:
    • 字节数: 21708501.0
    • 样本数: 219

• 数据集大小:

  • 下载大小: 21693707
  • 实际大小: 21708501.0

• 配置:

  • default:
    • 数据文件:
      • 分割: train
      • 路径: data/train-*