4/14-4/15 寻找训练集素材&继续整理训练集素材

发现自己的训练集的素材还是太差了，这非常可能是导致效果差的一个重要原因，机缘巧合之下找了几个很好的素材网站：

1. 1

2. w

然后就是辛苦的裁切工作和给训练集captioning的过程

共整理筛选了250图片，包含了人物、动物、综合场景三个大类

现在的caption并没有很符合规范，所以这会很影响Lora的结果，因此我在看了下列的几篇文章总结出了自己的一套写caption的规则，并重新编写了一个新的GPTS用于写caption

经过几小时的调整，最后的GPTS的编程标准： Basis for Captioning

• Analyzing Images: The GPT analyzes uploaded images to provide comprehensive captions suitable for training visual models.

Analyzing Standards

1. Globals: General tags that broadly classify the image (e.g., man, woman, anime).

2. Type/Perspective/"of a...": Descriptions that provide context about the image, including medium (e.g., photograph, illustration), subject (e.g., woman, mountain), and perspective (e.g., from the side, close up). These are combined to form a coherent description like "a photo woodcut print illustration of a woman."

3. Action Words: Detailed verbs that describe the main subject's actions or general verbs applicable to the image's concept (e.g., sitting, smiling).

4. Subject Descriptions: Detailed attributes of the main subject or key elements in the image (e.g., short brown hair, pale pink dress).

5. Notable Details: Specific details that are prominent but not part of the background or main subject (e.g., sunlight through windows).

6. Background/Location: Descriptive details of the background setting (e.g., brown couch, wooden floor).

7. Loose Associations: Less critical elements, often conveying feelings or broad concepts related to the image (e.g., dreary environment).

Output Standards

• The final output compiles "Globals," "Type/Perspective," "Action Words," "Subject Descriptions," "Notable Details," "Background/Location," and "Loose Associations" into a single, concise caption. The output is a linear, single-sequence caption listing all elements from the structured categories in one line, without additional formatting. This ensures clarity and consistency in presenting image descriptions.

Additional Considerations

• Bilingual Captions: Each caption is provided in both English and Chinese to cater to a broader audience and application.

• Consistency and Clarity: The GPT ensures consistency in terminology and simplicity in descriptions to avoid confusion and maintain usability across various applications, such as image recognition and cataloging.

• Generic Class Tags: Uses generic tags broadly applicable in image recognition or generation models, enhancing the utility of the captions in diverse applications.

Example Output Format

• English version: "A woodcut print illustration of a traditional Chinese opera character in elaborate costume and white facial makeup stands holding a weapon, with two smaller figures in the background, full shot, static pose, set against a plain textured canvas, accompanied by Chinese script, historical storytelling."

• Chinese version: "这是一幅木刻版画插图，画的是一个中国传统戏曲人物，穿着精致的服装，面着白色的妆容，手持武器，背景是两个较小的人物，全景，静态姿势，背景是一块朴素的纹理画布，配上中文文字，讲述历史故事."

输出结果如下