hybrid-dit-176m

A 176M-parameter text-to-image latent diffusion model at 256×256, trained with Dew.

Try it in your browser: the hybrid-dit-176m Space runs this model on a shared CPU host, from an editable cell.

Architecture	Dew's hybrid_dit: 16 blocks of width 768, 12 heads, patch size 2; S5 state-space blocks and attention blocks in a 3:1 ratio, zigzag scan order, 2D spatial fusion
Parameters	175,640,848 (the denoiser; the text encoder and VAE are separate)
Resolution	256×256, denoised as 32×32×4 latents
Text conditioning	CLIP ViT-L/14 (openai/clip-vit-large-patch14) text encoder, frozen
Autoencoder	Stable Diffusion VAE (pcuenq/sd-vae-ft-mse-flax), frozen
Diffusion	EDM parameterization
Training	1,350,000 steps at batch size 128 on laion12m_coco: LAION-Aesthetics 12M (aesthetic score ≥ 6) and MS-COCO 2017
Weights	The exponential moving average of the training weights (decay 0.999)

Use

Install Dew from GitHub at a commit this model is tested with (add the [cuda13] extra on an NVIDIA GPU):

pip install "dewml @ git+https://github.com/AshishKumar4/dew@e51ace7452e0956aa15e95bec33e7011d6327690"

Then sample:

from dew.sampling import CFG, DPMSolverMultistep, TextToImage

pipe = TextToImage.from_pretrained("dewml/hybrid-dit-176m")
images = pipe(["a red fox in a snowy forest"], key=0, steps=20,
              solver=DPMSolverMultistep(), guidance=CFG(5.0)).host().images

images is a float array in [-1, 1] of shape [prompts, 256, 256, 3]; dew.artifacts.uint8_pixels converts it to 8-bit pixels.

Files

This is a Dew run directory: run.json records the model, the text encoder, the VAE and the training configuration, and 1350000/ is the checkpoint TextToImage.from_run and from_pretrained read. The checkpoint holds the weights, including the text encoder's and the VAE's, and no optimizer state: it samples, and does not resume training.

Limitations

The model draws at 256×256. It learned from web images and their alt text, with the biases that data has, and it sometimes draws the flat white or black bands of letterboxed images.