Text & analysis
30 text models that read an image or video and return text — captioning, OCR, classification, Q&A, and summarization. Unlike every other mode, these LLMs analyze media instead of generating it.
Quick start
They run through gen-ai describe (not generate), and print the answer to stdout:
bash
# describe an image (Claude Sonnet is the default)
gen-ai describe -i photo.jpg
# ask a specific question
gen-ai describe -i receipt.jpg -p "extract the total and tax"
# summarize a video (auto-routes to a video-capable model)
gen-ai describe --video clip.mp4 -p "summarize what happens"From an MCP client, the same models are reachable through picsart_generate with an imageUrls (or videoUrl) input.
Input types
| Type | Meaning | Models |
|---|---|---|
i2t | image → text | Claude Fable / Opus / Sonnet / Haiku, GPT, Gemini Flash |
v2t | video (or image) → text | Gemini 3 Pro |
Only Gemini 3 Pro accepts video, so gen-ai describe --video … auto-selects it unless you force another model with -m.
Providers
| Provider | Models | Highlights |
|---|---|---|
| Anthropic | Claude Fable, Opus, Sonnet, Haiku | Opus for hard reasoning; Haiku for high-volume |
| OpenAI | GPT-6 Astra, GPT-5.6 family, GPT-5 and GPT-4 families | Strong general image understanding |
| Gemini 3 Pro, Gemini 3.8 / 3.7 / 3.6 Flash, 3.5 Flash Lite, 2.5 Flash | Pro reads video; Flash tiers handle image analysis |
Common parameters
| Param | CLI flag | Notes |
|---|---|---|
prompt | -p | The question or instruction (optional — defaults to "describe this") |
imageUrls | -i | Image(s) to analyze |
videoUrl | --video | Video to analyze (Gemini 3 Pro only) |
thinking | --thinking | Reasoning depth, where the model supports it |
These models return text, so the CLI prints the answer to stdout and skips download / Drive save. Add -q to drop the model/time header (printed to stderr), or --json for { text, model, durationMs }.