Image Generation

Akio supports local text-to-image generation using the image (or img) command. It runs entirely on-device — no external API required.

Supported Models

Alias	HuggingFace Repo	Default Steps	Notes
Z-Image-Turbo	Tongyi-MAI/Z-Image-Turbo	9	Optimized for fast inference

Quickstart

1. Pull the model

akio pull Tongyi-MAI/Z-Image-Turbo

2. Generate an image

akio image --model turbo --prompt "a cute orange cat sitting on a rooftop looks at a sunset in studio ghibli-style"

The output is saved to z_image_output.png by default.

Options

Flag	Type	Default	Description
-m, --model	string	(required)	HuggingFace repo of the model to use
--prompt	string	(required)	Text prompt describing the image
--negative-prompt	string	""	Prompt to steer generation away from
--height	usize	model default	Output height in pixels (must be divisible by 16)
--width	usize	model default	Output width in pixels (must be divisible by 16)
--num-steps	usize	model default (9)	Number of denoising steps
--guidance-scale	f64	model default	Classifier-free guidance (CFG) scale
--seed	u64	random	Seed for reproducible outputs
--output	string	output.png	Output filename
--cpu	flag	false	Force CPU inference instead of GPU

Examples

Custom resolution and seed

akio img \
  --model turbo \
  --prompt "a futuristic city skyline, neon lights, cinematic" \
  --width 1024 \
  --height 576 \
  --seed 42 \
  --output city.png

With negative prompt and guidance scale

akio img \
  --model turbo \
  --prompt "a serene mountain lake at dawn, photorealistic" \
  --negative-prompt "blurry, low quality, distorted" \
  --guidance-scale 3.5 \
  --output lake.png

CPU-only inference

akio img \
  --model turbo \
  --prompt "a red panda in a bamboo forest" \
  --cpu \
  --output panda.png

Notes

Image dimensions

Width and height must both be divisible by 16. If you pass an invalid size, Akio will suggest the nearest valid dimensions.

Reproducibility

Pass --seed with a fixed integer to get the same image across runs. Without it, a random seed is used each time.

GPU acceleration

GPU is used by default when available (Metal on macOS, CUDA on Linux/Windows). Use --cpu to override.

Model weights location

Weights are stored in the same directory as models downloaded via akio pull.