More meaningful descriptions

[isaacbernat]

TLDR: The current prompt generates many superfluous words, the space is limited and therefore should be avoided.

There are only 1-line descriptions per element. That restriction severely limits the amount of words allowed. Therefore, emphasis should be made in concision and avoidance of space wasted by words with little meaning. The current prompt (version 0.3.0, which uses gemini-2.5-flash-preview-04-17) has many of these. E.g. "small", "large" and "pixel art" were in the top 25 most occurring words (3500, 2000 and 1000 times) within the first 600 images described. Words which refer to element locations or size are superfluous since we have X, Y coordinates. Feel free to contribute and improve the prompt.

By calling the following function from create_description_page one may get a rough estimate of the words to optimise/reduce (more frequent ones will have bigger impact):

def common_words(descriptions):
    word_counts = Counter()
    for day_num, current_description in enumerate(descriptions.values()):
        for entry in current_description:
            text_lower = entry.lower()
            words = re.findall(r'\b\w+\b', text_lower)
            word_counts.update(words)
    return word_counts.most_common()

The current prompt is located in analyze_image_with_metadata function. The current description may be found here

[isaacbernat]

I'll merge in a few days an update which uses gemini-3-flash-preview, so the descriptions should improve. That being said, I don't plan on updating the existing CSV, among other reasons because RPD (Requests Per Day) on the free tier for this model is limited to 20, and we have 800+ images already... Feel free to update if you have access to a better model and/or one with a bigger quota.