HomeIMG-1.2B
Polish and English image generator
1 min readupdated
A text-to-image model based on DiT (Diffusion Transformer) with 1.24 billion parameters and a trainable text refiner. It uses EMA (Exponential Moving Average) weights in safetensors format and a frozen Poziomka-Baza encoder. The solution was trained on images ranging from 256 to 1024 pixels using FLUX.2 [klein] base 4B as a teacher. The model is suitable for researchers and developers interested in experimenting with bilingual image generation.
Key features
- DiT diffusion architecture with 1.24B parameters
- Trainable text refiner with 72.1M parameters
- Support for prompts in Polish and English
- EMA weights in safetensors format
What you can do
- Generate images at resolutions up to 1024x1024 pixels based on a text description
- Test generation quality across different step counts and time shifts
- Run the model using the diffusers Python library
- Compare model behavior using various generator seeds
Example
import torch
from diffusers import DiffusionPipeline
REPO = "cpral/HomeIMG-1.2B-PL-EN-epoch49.2-DanceOPD3652"
pipe = DiffusionPipeline.from_pretrained(
REPO, custom_pipeline=REPO, trust_remote_code=True
).to("cuda")
image = pipe(
"rudy lis siedzący w zaśnieżonym lesie o wschodzie słońca, fotografia",
resolution=1024,
num_inference_steps=50,
guidance_scale=3.0,
time_shift=6.0,
generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("lis.png")Read more
DetailsArchitecture: DiT 1.24B, 72.1M refiner
Training: 1604 updates on DanceOPD2048 base
Resolution: 256, 512, and 1024 px
License: Apache-2.0, openmdw-1.1 for Poziomka










Comments