Capybara for Text to Image

Capybara is a unified visual generation model that can do text‑to‑image, image editing, and video tasks, but here you’d use it mainly for text‑to‑image to create high‑quality still images from prompts.​

What it is

• A 14B diffusion‑transformer model (built on HunyuanVideo 1.5) that supports T2I, T2V, I2I, and V2V in one architecture, with custom ComfyUI nodes.

• For text‑to‑image, you give a natural‑language prompt and it generates 720p‑class images with strong realism and style flexibility.

Key features (text to image)

• Handles complex scenes (multiple characters, detailed environments) while keeping good global composition.​

• Supports instruction‑like prompts (“cinematic close‑up,” “anime style,” “studio product shot”) thanks to its unified semantic/vision transformer design.​​

• Recommended settings around 720p, ~50 steps for best quality, with the option to reduce steps using acceleration LoRAs for faster renders.​​

• Tight ComfyUI integration via official templates like “Capybara: Text to Image,” so you can drop it into existing node graphs easily.​

Best use cases

• Cinematic keyframes and concept art from detailed text briefs (characters, lighting, camera language).​

• Stylized or realistic illustrations for thumbnails, posters, and social content when you don’t need separate models for video.​

• Unified pipelines where you might later extend a still image into motion (I2V/T2V) using the same Capybara model family.