Vision Service: Improve caption prompt and allow submitting a custom prompt with the request data

[lastzero]

This task includes improving the default caption generation prompt(s) and allowing a custom prompt to be submitted:

• For the intended use, image captions should not start with "Here's a detailed description of the image:...", "This image...", or "An image of...", but should simply describe the content of the image (and not be extremely long or markdown formatted, see screenshot below).
• In addition, it should be possible to submit a custom prompt in the request data along with the model name and version, so that users can customize the prompt for the model they have chosen without touching the photoprism-vision service.

Examples of good/better image captions:

• A cat sleeping with its head resting on the strings of an instrument.
• A vibrant, full poppy flower in rich shades of red and pink.
• A young woman, likely in her late 20s or early 30s. She is smiling broadly, appearing friendly and approachable.

Of course, the caption doesn't have to start with "A" and can be longer than these examples.

To illustrate, here is a screenshot of the captions generated by the gemma3 model with the current prompt:

Related Issues:

• Vision Service: Use an inference server or framework instead of custom microservices #1 (comment)

[lastzero]

Llama 3.2 Vision

Besides gemma3 and minicpm-v, Llama 3.2 Vision seems to be another Ollama Model that might be worth testing/using for caption generation:

• https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/
• https://ollama.com/library/llama3.2-vision
• https://ollama.com/search?c=vision

The version with 11B parameters has a size of 7.9GB, so it works on mid-range NVIDIA GPUs with 8GB of RAM:

Model	Parameters	Size	Download
Llama 4	109B	67GB	ollama run llama4:scout
Llama 4	400B	245GB	ollama run llama4:maverick
Llama 3.3	70B	43GB	ollama run llama3.3
Llama 3.2	3B	2.0GB	ollama run llama3.2
Llama 3.2	1B	1.3GB	ollama run llama3.2:1b
Llama 3.2 Vision	11B	7.9GB	ollama run llama3.2-vision
Llama 3.2 Vision	90B	55GB	ollama run llama3.2-vision:90b

Here's a demo video showing its capabilities:

AQN57z1yVWL9PAN8EmBkBwlZ_QO1xLJZSxZ-p0YEXVwnqATTSj45aPDxuMUsMIzgXDgVpJYtEjR36SHUmXixKP2F.mp4

[lastzero]

With the new default prompt and the Llama 3.2 Vision model, captions have improved a lot and sound more natural:

However, it seems that the number of subjects is not well recognized or there's a problem with how the image is passed to the model (if multiple thumbnails are passed to the model, that could possibly cause this):