A quick test of OpenRouter LLM, a ComfyUI workflow that sends chat requests to a curated list of Large Language Models. In short, it asks each LLM to write a prompt for an input image. There are 21 LLMs to chose from. I tested four of them. The top image (A) is a news photo from Getty Images. (B) Claude Opus 4.7: military quadruped robot dog in foreground, motion blur, glowing blue and green sensor lights, olive drab tactical paint, walking through tall grass, operator in background holding remote controller tablet, masked face with neck gaiter, black baseball cap, navy long-sleeve tactical shirt, tan cargo pants, seated on fallen log, birch and deciduous forest backdrop, overcast dusk lighting, muted green palette, shallow depth of field, documentary photojournalism style, cinematic realism, 35mm lens, natural ambient light, gritty defense technology field test atmosphere (C) Qwen 6.3 max preview: Action shot of a quadruped military robot dog leaping towards the camera in a forest, motion blur, shallow depth of field, blue glowing sensor eye. In the background, a tactical operator wearing a black balaclava, cap, and cargo pants holds a remote controller. Forest clearing, natural overcast lighting, realistic photography style. (D) Google/Gemini 3.4 Flash: realistic photograph, metallic gray quadruped robot mid-leap toward camera, detailed mechanical joints and black rubber feet, glowing blue circular light and green sensor on head, motion blur on legs, low angle shot, dense green forest background with trees and grass, man in navy blue t-shirt, tan cargo pants and black balaclava holding black remote controller, standing near fallen logs, natural overcast daylight, shallow depth of field (E) x-AI Grok 4.3: A tactical quadrupedal robot dog leaping mid-air towards the camera in a forest, shallow depth of field, blurred motion foreground. In the background, a man wearing a balaclava, dark blue shirt, and cargo pants holding a remote controller. Natural daylight, dense woods, dry grass, realistic action shot, muted color palette. They also had a parameter labeled 'Reasoning_Effort' with options Off, Low, Medium and High. On most models you can input up to 20 images and also videos, so this just scratches the surface.