sana_bullet

Fast finetuned one-step Sana checkpoint for the standalone sana_bullet GPU inference stack.

Links

• GitHub repo: rotem154154/sana_bullet
• Model repo: irotem98/sana_bullet

One-Line Generation

# pip install "git+https://github.com/rotem154154/sana_bullet.git"
from sana_bullet import generate_image

image = generate_image(
    "a cinematic portrait of a fox in soft winter light",
    model="irotem98/sana_bullet",
)
image

from_pretrained(...) Loading

# pip install "git+https://github.com/rotem154154/sana_bullet.git"
from sana_bullet import SanaBulletPipeline

pipe = SanaBulletPipeline.from_pretrained("irotem98/sana_bullet")
image = pipe.generate("a retro-futuristic motorcycle speeding through neon rain")
image

Server Usage

pip install "git+https://github.com/rotem154154/sana_bullet.git"
export SANA_BULLET_CHECKPOINT_REPO_ID="irotem98/sana_bullet"
python -m sana_bullet.ui_server

Files

• step0003000.pt: finetuned transformer checkpoint

Base Stack

• Base model: Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers
• VAE: dc-ai/dc-ae-lite-f32c32-diffusers

Notes

• The checkpoint is designed for GPU inference.
• The higher-level Python API lives in the GitHub repo, while this model repo only hosts the weights and model card.