Sana-Sprint¶

Distilled 1-2 step inference. ~0.1s per 1024px image on H100, ~0.3s on RTX 4090.

from strands_sana import sana_sprint_generate

sana_sprint_generate(
    prompt="a cyberpunk hummingbird",
    model="sana-sprint-1.6b-1024",   # or sana-sprint-0.6b-1024
    steps=2,                          # sweet spot
    seed=42,
)

How it works¶

Sprint uses sCM-LADD distillation:

sCM — continuous-time consistency distillation
LADD — Latent Adversarial Diffusion Distillation

The student model is trained to take giant denoising steps. Result: 1-2 steps gets the same quality as 20 steps of the teacher.

Step count¶

Sprint accepts steps ∈ {1, 2, 4, 8, ...}.

steps	speed	quality
1	fastest	sometimes blurry
2	default	best balance
4	slower	slight quality bump
8	slowest	diminishing returns

Implementation note

Sprint's intermediate_timesteps=1.3 default is only valid for steps=2. The wrapper passes intermediate_timesteps=None automatically when steps≠2. (Caught by BUG #5.)

Sprint Img2Img¶

Sprint's image-to-image variant uses the same checkpoint:

from strands_sana import sana_img2img

sana_img2img(
    prompt="anime cel-shaded style",
    image_path="src.png",
    model="sana-sprint-i2i-1.6b-1024",
    strength=0.5, steps=2,
)

→ See sana_img2img.

Demo¶

Same prompt, same seed, two models: