Skip to content

Sana

Model Source Revision Weight license Parity (CPU, fp32)
sana-1600m-1024 Efficient-Large-Model/Sana_1600M_1024px_diffusers ac0da2f Apache-2.0 Verified
sana-600m-1024 Efficient-Large-Model/Sana_600M_1024px_diffusers 28f3af7 Apache-2.0 Verified
sana-sprint-1.6b-1024 Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers 19683c5 Apache-2.0 Verified
sana-1.5-4.8b-1024 Efficient-Large-Model/SANA1.5_4.8B_1024px_diffusers 9468102 Apache-2.0 Verified

Catalog: sana · Models: sana-1600m-1024, sana-600m-1024, sana-sprint-1.6b-1024, sana-1.5-4.8b-1024 · Task: t2i

Architecture. A Gemma-2 2B text model (26 layers, width 2304, soft-capped attention and alternating sliding-window layers) encodes the prompt, and its last hidden state conditions a linear diffusion transformer. The transformer applies ReLU linear attention over the image tokens, cross-attention to the masked text, GLUMBConv feed-forward blocks and AdaLN-single timestep modulation. The 1600M model has 20 blocks of width 2240 and the 600M model has 28 blocks of width 1152. The 1.5 4.8B model uses 60 blocks of width 2240 with qk normalization. The Sprint 1.6B model keeps the 1600M width but replaces the timestep embedding with a combined timestep + guidance embedding and uses the single-step consistent-model (SCM) scheduler with trigflow parameterization. Images are decoded by the DC-AE autoencoder, which compresses 32× into 32 latent channels (scaling factor 0.41407). The standard models use the order-2 midpoint DPM-Solver++ with flow sigmas and flow_shift 3.0; Sprint uses the 2-step SCM schedule. Before encoding, a "complex human instruction" is prepended to the prompt, and the first token plus the last 299 tokens of the encoder output are kept.

Usage.

python -m imagehub.pretrained sana-1600m-1024 checkpoints/sana-1600m-1024
from imagehub import pipeline

with pipeline(
    "t2i", backend="sana-1600m-1024", model="checkpoints/sana-1600m-1024", device="mps"
) as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. All models default to guidance 4.5 and 1024×1024, with an empty negative prompt. sana-1600m-1024, sana-600m-1024 and sana-1.5-4.8b-1024 default to 20 steps; the standard options are seed, steps, guidance, width, height, negative_prompt, use_resolution_binning, max_sequence_length (default 300) and complex_human_instruction. sana-sprint-1.6b-1024 defaults to 2 steps, ignores negative_prompt (guidance is embedded via the transformer) and supports the same remaining options. With use_resolution_binning on (the default), the requested size is snapped to the closest aspect-ratio bin, the image is generated at the bin size, and the decoded image is resized and centre-cropped back to the request; turn it off to generate at the exact size, which must then be a multiple of 32. complex_human_instruction is a list of strings; None selects the published instruction and an empty list disables it. The instruction is never added to the negative prompt. Guidance 1.0 or lower skips the unconditional pass for standard models. Unrecognized settings raise an error.

Memory. In float32 the 1600M transformer needs about 6.4 GB (2.4 GB for the 600M model), the Gemma-2 text model about 10.5 GB and DC-AE about 1.3 GB. The 1.5 4.8B and Sprint transformers need roughly 9–10 GB. The Gemma-2 weights are stored in bfloat16 and upcast on load.

Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.

Out of scope. The 2K and 4K checkpoints, ControlNet, PAG, fp16/bf16 parity and caption cleaning are not supported. Caption cleaning requires external packages; parity is against clean_caption=False. Only decoding is exposed for DC-AE.

License. All repositories declare the Apache-2.0 license. The text encoder is a Gemma-2 checkpoint; check Google's Gemma terms of use before redistributing it. imagehub does not distribute weights.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.