arraypress/imagine-sana-sprint-1.6b

SANA Sprint 1.6B — Core AI

Fast pictures from words, on Apple silicon: NVIDIA's SANA Sprint 1.6B converted to a Core AI diffusion bundle, 1024 × 1024 in 2 steps — about 1.8 s a picture on an M3 Max.

Credit and licences — three parts, three licences

Read NOTICE first.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

part	from	licence
Transformer.aimodel	SANA Sprint 1.6B, NVIDIA	Apache-2.0 — LICENSE
VAEDecoder.aimodel	DC-AE, MIT HAN Lab	MIT — LICENSE-DC-AE
TextEncoder.aimodel, tokenizer/	Gemma 2 2B IT, Google	Gemma Terms of Use — GEMMA-TERMS.txt

The text encoder may not be used for anything in the Gemma Prohibited Use Policy (GEMMA-PROHIBITED-USE-POLICY.txt), and anyone redistributing these files must pass that restriction on. All parts allow commercial use.

Conversion: Apple's official diffusion recipe, apple/coreai-models (BSD-3) at commit 52dbd93, in bfloat16 without quantization. The recipe lists only the 0.6B model; one line was added to its registry (python/src/coreai_models/diffusion/models.py):

("sana-sprint-1.6b", "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers", "sana_sprint"),

The exporter reads every size from the model's own config, so nothing else changed. These files are therefore modified (converted) versions of the originals. This repository only converts and hosts the weights; all credit belongs to NVIDIA, MIT HAN Lab and Google.

Shapes

shape	where	extra download	extra compiled cache
1:1, 1024 × 1024	top level	—	—
16:9, 1344 × 768	aspect-16x9/ — transformer and DC-AE decoder traced at 1344 × 768 (42 × 24 latent); shares ../TextEncoder.aimodel	3.3 GB	~3.3 GB
9:16, 768 × 1344	aspect-9x16/ — transformer and DC-AE decoder traced at 768 × 1344 (24 × 42 latent); shares ../TextEncoder.aimodel	3.3 GB	~3.3 GB
4:5, 896 × 1120	aspect-4x5/ — transformer and DC-AE decoder traced at 896 × 1120 (28 × 35 latent); shares ../TextEncoder.aimodel	3.3 GB	~3.3 GB

9:16 and 4:5 (added 2026-10-13) were exported with COREAI_IMAGE_HW=HEIGHTxWIDTH from the arraypress fork of apple/coreai-models at b667888 (tag 1.0.0-nonsquare.3), bfloat16, no compression, transformer and decoder only. Measured on an M3 Max through imagine --fast, warm: about 9–12 s for a one-picture command, the same seed giving a byte-identical file.

16:9 runs at about 4.0 s a picture once loaded (2.4 s for 1:1 in the same session). It needs the non-square support in arraypress/coreai-models, a fork of Apple's 1.0.0; 1:1 runs on Apple's pipeline as is. The 16:9 files are converted (modified) versions of the same three licensed parts and carry the same licences and notices as the rest of this repository.

Measured (macOS 27, M3 Max, 36 GB, GPU, 1024 × 1024)

Measure	Result
per picture, model loaded	1.83 s
first picture in a fresh process	28.7 s
peak memory	3.0 GB
disk	8.2 GB (4.9 GB of it the Gemma text encoder)
same seed, run to run	byte-identical

Of five candidates on the same eight prompts it was the fastest by far; unlike the 0.6B model, which wrote "Apcn" for "OPEN", it spelled a short sign readably ("Open", mixed case). It cannot edit pictures, and it does not fit an iPhone's per-app memory — the text encoder alone is 4.9 GB.

Use

With the imagine command-line tool, which fetches this on first use:

imagine --fast "a lighthouse at dusk, oil painting" -o draft.png