Generate title, caption and keywords from attached image using Gemini

gemini_describe.lua

#!/usr/bin/env lua
local http = require("socket.http")
local ltn12 = require("ltn12")
local mime = require("mime")  -- For base64
local json = require("dkjson")  -- dkjson is a JSON library for Lua

-- Load and encode image
local function read_image_base64(path)
  local file = assert(io.open(path, "rb"))
  local content = file:read("*all")
  file:close()
  return mime.b64(content)
end

-- JSON prompt message
local prompt_text = [[
You are a professional photography analyst with expertise in object recognition and computer-generated image description. 
You also try to identify famous buildings and landmarks as well as the location where the photo was taken. 
Furthermore, you aim to specify animal and plant species as accurately as possible. Always give common name followed by the scientific name in brackets e.g (Beta vulgaris).
You also describe objects—such as vehicle types and manufacturers—as specifically as you can.

Analyze the uploaded photo and generate the following data:
* Keywords (comma-separated list of 50 single-word keywords)
* Image title
* Image caption (Maximum 200 characters)

Make sure the result is in JSON format:
{
"title": "",
"caption": "",
"keywords": "key1,key2"
}
]]

-- Setup request payload
local function build_payload(image_base64)
  return json.encode({
    contents = {
      {
        role = "user",
        parts = {
          { text = prompt_text },
          {
            inlineData = {
              mimeType = "image/jpeg",
              data = image_base64
            }
          }
        }
      }
    },
    generationConfig = {
      temperature = 1,
      topP = 0.95,
      maxOutputTokens = 8192,
      responseModalities = { "TEXT" }
    },
    safetySettings = {
      { category = "HARM_CATEGORY_HATE_SPEECH", threshold = "BLOCK_NONE" },
      { category = "HARM_CATEGORY_DANGEROUS_CONTENT", threshold = "BLOCK_NONE" },
      { category = "HARM_CATEGORY_SEXUALLY_EXPLICIT", threshold = "BLOCK_NONE" },
      { category = "HARM_CATEGORY_HARASSMENT", threshold = "BLOCK_NONE" }
    }
  })
end

-- Make API call
local function generate()
  local image_base64 = read_image_base64("20250506-153543.jpg")
  --print("Image: " .. image_base64 .. "\n")
  local payload = build_payload(image_base64)
  --print("Payload: " .. payload .. "\n")

  local response = {}
  local _, status = http.request{
    url = "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash-001:streamGenerateContent?key={your_api_key}",
    method = "POST",
    headers = {
      ["Content-Type"] = "application/json",
      ["Content-Length"] = tostring(#payload)
    },
    source = ltn12.source.string(payload),
    sink = ltn12.sink.table(response)
  }

  if status == 200 then
    --print(table.concat(response))
    local decoded, _, err = json.decode(table.concat(response))

    if not decoded then
      print("JSON decode error:", err)
      return
    end

    -- Combine all `parts[].text` from every item in the array
    local combined_text = ""

    for _, item in ipairs(decoded) do
      if item.candidates and item.candidates[1] and item.candidates[1].content and item.candidates[1].content.parts then
        local parts = item.candidates[1].content.parts
        for _, part in ipairs(parts) do
          if part.text then
            combined_text = combined_text .. part.text
          end
        end
      end
    end
    
    combined_text = string.gsub(combined_text, '```json', '')
    combined_text = string.gsub(combined_text, '```', '')

    print("Combined output:\n", combined_text)
  else
    print("Request failed with status:", status)
  end
end




generate()