Images in 8 steps
Core model parameters 7B generation + 8B encoder
Model files 9.51 GB
A poster from one sentence, with legible Chinese and English text, in 7 aspect ratios, all on the device
- 1.5K detail
- Chinese & English text
- 7 aspect ratios
Recorded on the board
Screen recording of the board web demo, recorded on Qualcomm QCS8550. (waiting segments sped up)
Highlights
Eight steps per image
A built-in 8-step sampling schedule, nothing to tune.
Readable text in the picture
Titles, signs and labels come out as written, in Chinese and English.
Long prompts in Chinese and English
Short or very detailed: up to 4096 tokens at 512 and 1K, several thousand characters.
7 aspect ratios, 3 sizes
1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16 at 512, 1K and 1.5K, up to 1792 × 1344.
Fully on-device
The text encoder, the diffusion model and the image decoder all run on the board's NPU.
Easy to integrate
A small C API and a Python binding; the same prompt, size and seed give the same image.
Pipeline
From input to output: where every step runs.
- 01Browser
Prompt in
A description in Chinese or English, with size, aspect ratio and seed
- 02NPU
Text encoder
Encodes the prompt
- 03NPU
Diffusion
The diffusion model, 8 denoising steps
- 04NPU
Image decoder
The VAE decodes the image tile by tile
- 05Browser
Image out
The image with the time of each stage; download as PNG
Supported hardware
- Core model parameters
- 7B generation + 8B encoder
- Model files
- 9.51 GB
- License
- Pandora License