Sana 1600M 4Kpx BF16
The 4K checkpoint of NVIDIA's 1.6B Sana model, fine-tuned from the 1024px BF16 release to generate images up to 4096 × 4096, with prompts in English, Chinese or emoji.
Overview
Sana 1600M 4Kpx BF16 is the largest-canvas member of the 1.6B Sana family. NVIDIA released it in January 2025, fine-tuned from the 1024px BF16 checkpoint. The card gives its target as 4Kpx-based images in varying heights and widths, and NVIDIA's own example renders 4096 × 4096.
The parameter count does not change with resolution: 1648M in a Linear Diffusion Transformer, with Gemma2-2B-IT encoding text and DC-AE handling the pixel space. DC-AE compresses each side 32 times, so a 4096 × 4096 image works out to a 128 × 128 latent, the same latent size an 8x autoencoder would give a 1024 × 1024 image.
Choosing a variant: use this one when the deliverable must be 4K, or wider than 2K on a side. Sana 1600M 2Kpx BF16 targets 2K output, while Sana 1600M 1024px BF16 and its FP16 counterpart Sana 1600M 1024px target 1024px. Sana Text to Image puts all four checkpoints behind one model picker.
What you can build
- Large prints, billboards and display backgrounds that need 4096 pixels across.
- Interior and architecture renders, like the luxury home among the card's 4K samples.
- Surreal art. The card's gallery includes a melting apple and a portrait set into broken porcelain.
- Master images for cropping, where one 4K render yields several smaller frames.
Inputs
Prompt and negative prompt are required, and the negative is filled in as "bad, ugly, low quality". Advanced options add an image count (1 to 8), a seed, width and height (each capped at 4096), steps (default 25, maximum 500) and guidance scale (default 3.5, between 1 and 20).
Outputs
You get the requested number of images at the requested size. The page samples are 1024 × 1024 files, so they do not show this checkpoint at full resolution.
Recommended settings
- Width and height both start at 1024 here. Set them to 4096, or to another 4K-class size for a different aspect ratio, to use the resolution this checkpoint was fine-tuned for.
- NVIDIA's example pairs 4096 × 4096 with guidance scale 5.0 and 20 steps.
Limitations
NVIDIA's card lists the same gaps for the 4K model as for its siblings. It does not reach perfect photorealism and cannot render complex legible text. Fingers may be malformed. Complex scenes suffer from limited data, and the DC-AE autoencoder loses some information when encoding.
Safety & compliance
Sana's 4K weights carry the Apache 2.0 license, and the Gemma-2-2B-IT encoder is bound by Google's Gemma Terms of Use, along with its Prohibited Use Policy. NVIDIA's card says the model is intended for research purposes only, and names art, design, educational tools and generative-model research as research areas. Depicting real people or events as fact is out of scope, and the card cautions that image models can amplify social biases.
Example prompts
Great starting points for Sana_1600M_4Kpx_BF16_diffusers.
API quick start
Run Sana_1600M_4Kpx_BF16_diffusers with a single API call.
{
"prompt": "Cinematic photo, black panther, trending …",
"negativePrompt": "bad, ugly, low quality",
"steps": 25,
"scale": 3.5
}curl -X POST "https://api.wiro.ai/v1/Run/efficient-large-model/sana-1600m-4kpx-bf16-diffusers" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Cinematic photo, black panther, trending …",
"negativePrompt": "bad, ugly, low quality",
"steps": 25,
"scale": 3.5
}
JSONOther models you might like
Explore similar image generation models.
