Catrex 1.0 Image
Text-to-image model for face generation, fine-tuned on 20,000 captioned portraits.
Results
| Metric | Value |
|---|---|
| MSE loss | 0.1678 |
| Resolution | 192x192 |
| UNet parameters | 323M |
| Training steps | 500 |
| Base model | segmind/tiny-sd |
Python (diffusers)
from diffusers import StableDiffusionPipeline
import torch
pipe = StableDiffusionPipeline.from_pretrained(
"Catniti/catrex-1.0-image", torch_dtype=torch.float16, safety_checker=None
).to("cuda")
image = pipe(
"a photo of a young woman with long blonde hair, smiling",
num_inference_steps=30,
guidance_scale=7.5,
height=192, width=192,
).images[0]
image.save("face.png")
Prompt format
Plain English descriptions of a person: gender, age, hair, expression.
a photo of an older man with a grey beard and glasses
Limitations
Trained only on portraits, so it generates faces regardless of the prompt — ask for a landscape and you still get a face. Best results at 192x192; higher resolutions degrade.
- Downloads last month
- 22