Sana 1.5
NVIDIA's linear diffusion transformer with a Gemma 2 text encoder and a 32x autoencoder, in 1.6B and 4.8B sizes. Makes 1024x1024 images in 20 steps.
Conditional for size 1.6b: code Apache-2.0, weights Apache-2.0. Measured on our RTX 4090: 1.53 s per 1024×1024 image at 20 steps, 12.1 GiB VRAM peak. We scored it 0.81 on GenEval. safetensors weights, 9.02 GB.
Conditional for size 4.8b: code Apache-2.0, weights Apache-2.0. Measured on our RTX 4090: 4.14 s per 1024×1024 image at 20 steps, 18.0 GiB VRAM peak. We scored it 0.80 on GenEval. safetensors weights, 14.83 GB.
License
Conditional
Conditional
- Code
- Apache-2.0checked 11 Nov 2026
- Weights
- Apache-2.0all sizes, checked 11 Nov 2026
Apache-2.0 for Sana's own weights. The repository also holds its text encoder, Gemma 2 2B IT, which the model card puts under the Gemma Terms of Use: commercial use is allowed, under the Gemma Prohibited Use Policy.
Apache-2.0 for Sana's own weights. The repository also holds its text encoder, Gemma 2 2B IT, which the model card puts under the Gemma Terms of Use: commercial use is allowed, under the Gemma Prohibited Use Policy.
On our RTX 4090, measured
1.53s per image
p90 1.54 s, bf16, batch 1, 1024x1024, 20 steps, guidance 4.5
- VRAM peak
- 12.1 GiB
- Status
- Runs as is
Environment, 11 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, diffusers==0.42.1. DPMSolverMultistepScheduler, seed 0 on cpu. Baseline 801 MiB. 2 warm-up + 10 timed images, median of 3 runs. Scope: predict(): one image from the prompt: text encoding, every denoising step and the VAE decode, to a PIL image; prompt read beforehand.
Accuracy
0.81GenEvalmeasured
GenEval, 2,212 images, bf16, GenEval af4902f · 11 Nov 2026
Reported0.82 GenEval overall on GenEvalSource: NVlabs/Sana README at commit b3bad95, performance table
4.14s per image
p90 4.15 s, bf16, batch 1, 1024x1024, 20 steps, guidance 4.5
- VRAM peak
- 18.0 GiB
- Status
- Runs as is
Environment, 11 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, diffusers==0.42.1. DPMSolverMultistepScheduler, seed 0 on cpu. Baseline 801 MiB. 2 warm-up + 10 timed images, median of 3 runs. Scope: predict(): one image from the prompt: text encoding, every denoising step and the VAE decode, to a PIL image; prompt read beforehand.
Reported4.2 s per imageA100, bf16, batch 1, 20 sampling stepsSource: paper (arXiv 2501.18427v4), table 1
Accuracy
0.80GenEvalmeasured
GenEval, 2,212 images, bf16, GenEval af4902f · 12 Nov 2026
Reported0.81 GenEval overall on GenEvalSource: paper (arXiv 2501.18427v4), table 1
Weights
safetensors
- Size
- 9.02 GB
- sha256
300b4109622d8ffd12d827a4127d2f82f102e3b9a0d413285f6bcf46729a3e54
safetensors
- Size
- 14.83 GB
- sha256
776004ceb0d1cd491ffad91896078b6e7d804a82b81cba262084ca0475ce92a2
Images for the shared prompts
Same prompts and seed for every model of this task, at each model's official settings. Failures are marked by hand.
14 of 14 prompts made, 8 marked as failures.

Counting
The prompt
Exactly five green apples in a single row on a wooden table, plain white background, product photo.
A correct image: Exactly five apples, all green, in one row on a wooden surface.

Color binding
Fails: The yellow shape is a pyramid on top of the cube, not a cone beside themThe prompt
A blue cube resting on top of a red sphere, with a yellow cone standing beside them, on a grey studio floor.
A correct image: Each color on its own shape: a blue cube on top of a red sphere, and a yellow cone beside them. No swapped or mixed colors.

Spatial relations
The prompt
A cat sitting on the left end of a sofa and a dog sitting on the right end, both facing the camera, living room.
A correct image: A cat on the left and a dog on the right, both on one sofa, both facing the viewer.

Text rendering
Fails: Reads FRESH BREAD DALYThe prompt
A wooden sign above a bakery door that reads "FRESH BREAD DAILY" in white painted capital letters.
A correct image: The sign reads FRESH BREAD DAILY, spelled exactly and legible, with no extra letters.

Text rendering, several lines
Fails: Garbled words and a wrong number instead of the three linesThe prompt
A café chalkboard menu with three lines written in chalk: "Coffee 3", "Tea 2" and "Cake 4".
A correct image: Three legible lines, each spelled exactly with its number: Coffee 3, Tea 2, Cake 4. No other words.

Photorealism, landscape
The prompt
Aerial photograph of a winding river through an autumn forest at sunrise, thin mist lying over the water.
A correct image: Reads as a photo from above: a winding river, autumn-colored trees, mist on the water, with no duplicated or melted trees.

Photorealism, food
Fails: Four or five slices of pork, not twoThe prompt
Close-up photo of a bowl of ramen with a soft-boiled egg cut in half, sliced green onions and two slices of pork, steam rising, on a dark wooden table.
A correct image: A ramen bowl with a halved soft-boiled egg, green onion and two slices of pork, on a dark wooden table.

Transparency and reflection
Fails: The water is full of bubbles, and no street shows through itThe prompt
A glass of water on a windowsill, a rainy city street outside the window, the street seen bent and upside down through the water in the glass.
A correct image: A clear glass of water on a sill, rain outside, and the street visibly refracted through the glass.

Hands
Fails: Six fingers on the right handThe prompt
Close-up of an adult's two hands kneading bread dough on a floured wooden counter, no face in the frame.
A correct image: Two hands with five fingers each and plausible joints, kneading dough on a floured counter. No face.

Fine detail
Fails: Shows about a quarter past ten, with subdials and no second hand at sixThe prompt
Macro photograph of a plain mechanical wristwatch with no brand name, its hands showing ten past ten, the second hand pointing at six.
A correct image: The hour and minute hands show 10:10 and the second hand points at 6. The dial has no brand or logo.

Long, dense prompt
Fails: Two cups but no teapotThe prompt
A cozy reading nook: a green armchair with a striped blanket over its arm, a sleeping orange cat curled up on the seat, a floor lamp with a yellow shade to the right of the chair, a small round table to the left holding a white teapot and two cups, a stack of three books on the floor, and rain running down the window behind.
A correct image: All of it: green armchair, striped blanket, orange cat asleep on the seat, floor lamp with a yellow shade on the right, round table on the left with a white teapot and two cups, three books stacked on the floor, rain on the window.

Style
The prompt
A watercolor painting of a lighthouse on a rocky coast under a stormy sky, loose brush strokes, visible paper texture.
A correct image: Looks like a watercolor, with washes and paper texture, not a photo: a lighthouse on rocks under a stormy sky.

Negation
The prompt
An empty park bench in fresh snow, with no people and no birds anywhere.
A correct image: A park bench in snow. No people and no birds anywhere in the image.

Unusual combination
The prompt
A giraffe wearing a long knitted red scarf, standing in a snowy city square at night, street lamps glowing.
A correct image: A giraffe with a red knitted scarf around its neck, in a snowy city square at night, with lit street lamps.
14 of 14 prompts made, 7 marked as failures.

Counting
The prompt
Exactly five green apples in a single row on a wooden table, plain white background, product photo.
A correct image: Exactly five apples, all green, in one row on a wooden surface.

Color binding
The prompt
A blue cube resting on top of a red sphere, with a yellow cone standing beside them, on a grey studio floor.
A correct image: Each color on its own shape: a blue cube on top of a red sphere, and a yellow cone beside them. No swapped or mixed colors.

Spatial relations
The prompt
A cat sitting on the left end of a sofa and a dog sitting on the right end, both facing the camera, living room.
A correct image: A cat on the left and a dog on the right, both on one sofa, both facing the viewer.

Text rendering
Fails: Reads Freshh Bread over DAY I DAYThe prompt
A wooden sign above a bakery door that reads "FRESH BREAD DAILY" in white painted capital letters.
A correct image: The sign reads FRESH BREAD DAILY, spelled exactly and legible, with no extra letters.

Text rendering, several lines
Fails: Garbled words and wrong numbers, with drawings of cupsThe prompt
A café chalkboard menu with three lines written in chalk: "Coffee 3", "Tea 2" and "Cake 4".
A correct image: Three legible lines, each spelled exactly with its number: Coffee 3, Tea 2, Cake 4. No other words.

Photorealism, landscape
The prompt
Aerial photograph of a winding river through an autumn forest at sunrise, thin mist lying over the water.
A correct image: Reads as a photo from above: a winding river, autumn-colored trees, mist on the water, with no duplicated or melted trees.

Photorealism, food
Fails: Four or five slices of pork, not twoThe prompt
Close-up photo of a bowl of ramen with a soft-boiled egg cut in half, sliced green onions and two slices of pork, steam rising, on a dark wooden table.
A correct image: A ramen bowl with a halved soft-boiled egg, green onion and two slices of pork, on a dark wooden table.

Transparency and reflection
Fails: The water is full of bubbles, and no street shows through itThe prompt
A glass of water on a windowsill, a rainy city street outside the window, the street seen bent and upside down through the water in the glass.
A correct image: A clear glass of water on a sill, rain outside, and the street visibly refracted through the glass.

Hands
Fails: Six fingers on the right handThe prompt
Close-up of an adult's two hands kneading bread dough on a floured wooden counter, no face in the frame.
A correct image: Two hands with five fingers each and plausible joints, kneading dough on a floured counter. No face.

Fine detail
Fails: Shows about a quarter past ten, with a subdial and no second hand at sixThe prompt
Macro photograph of a plain mechanical wristwatch with no brand name, its hands showing ten past ten, the second hand pointing at six.
A correct image: The hour and minute hands show 10:10 and the second hand points at 6. The dial has no brand or logo.

Long, dense prompt
Fails: The table with the teapot is on the right, and one cup sits on the booksThe prompt
A cozy reading nook: a green armchair with a striped blanket over its arm, a sleeping orange cat curled up on the seat, a floor lamp with a yellow shade to the right of the chair, a small round table to the left holding a white teapot and two cups, a stack of three books on the floor, and rain running down the window behind.
A correct image: All of it: green armchair, striped blanket, orange cat asleep on the seat, floor lamp with a yellow shade on the right, round table on the left with a white teapot and two cups, three books stacked on the floor, rain on the window.

Style
The prompt
A watercolor painting of a lighthouse on a rocky coast under a stormy sky, loose brush strokes, visible paper texture.
A correct image: Looks like a watercolor, with washes and paper texture, not a photo: a lighthouse on rocks under a stormy sky.

Negation
The prompt
An empty park bench in fresh snow, with no people and no birds anywhere.
A correct image: A park bench in snow. No people and no birds anywhere in the image.

Unusual combination
The prompt
A giraffe wearing a long knitted red scarf, standing in a snowy city square at night, street lamps glowing.
A correct image: A giraffe with a red knitted scarf around its neck, in a snowy city square at night, with lit street lamps.
Use it when
You want counts and positions followed at 1024x1024: both sizes drew exactly five apples and kept left and right apart, and 4.8B also put every color on its own shape.
Avoid it when
You need legible text or reliable hands: both sizes garbled a three-line menu, misspelled a sign and drew a hand with six fingers.
Sizes
| Size | Per image | VRAM | Accuracy | Weights | Runs | License |
|---|---|---|---|---|---|---|
| 1.6b | 1.53 s1024x1024 · 20 steps | 12.1 GiB | 0.81 GenEval, measured0.82 GenEval overall paper | 9.02 GB | As is | Conditional |
| 4.8b | 4.14 s1024x1024 · 20 steps | 18.0 GiB | 0.80 GenEval, measured0.81 GenEval overall paper | 14.83 GB | As is | Conditional |
Estimated on other GPUs
Estimate, not measured. Worked out from our RTX 4090 run and NVIDIA's spec sheets. Each kind of GPU work is scaled by that GPU's peak for it: tensor cores at the precision the kernels used, CUDA cores, or memory bandwidth. The CPU's part of each call is kept at what it took on our Intel Core i7-10700K. How we estimate
Where a call's time goes on our RTX 4090: tensor cores, BF16 57%, memory-bound 32%, CUDA cores, FP32 8%, CPU and kernel launches 4%.
GeForce (13)
| GPU | Fits | Per image |
|---|---|---|
| GeForce RTX 5090Blackwell, 32 GB GDDR7 | Yes | ≈ 1.1 s |
| GeForce RTX 5080Blackwell, 16 GB GDDR7 | Yes | ≈ 2.0 s |
| GeForce RTX 5070 TiBlackwell, 16 GB GDDR7 | Yes | ≈ 2.4–2.5 s |
| GeForce RTX 5070Blackwell, 12 GB GDDR7 | No, needs 12 GB | Doesn’t fit |
| GeForce RTX 5060 Ti 16GBBlackwell, 16 GB GDDR7 | Yes | ≈ 4.5–4.6 s |
| GeForce RTX 5060Blackwell, 8 GB GDDR7 | No, needs 12 GB | Doesn’t fit |
| GeForce RTX 4080Ada Lovelace, 16 GB GDDR6X | Yes | ≈ 2.4 s |
| GeForce RTX 4070 TiAda Lovelace, 12 GB GDDR6X | No, needs 12 GB | Doesn’t fit |
| GeForce RTX 4070Ada Lovelace, 12 GB GDDR6X | No, needs 12 GB | Doesn’t fit |
| GeForce RTX 3090Ampere, 24 GB GDDR6X | Yes | ≈ 2.9 s |
| GeForce RTX 3080Ampere, 10 GB GDDR6X | No, needs 12 GB | Doesn’t fit |
| GeForce RTX 3070 TiAmpere, 8 GB GDDR6X | No, needs 12 GB | Doesn’t fit |
| GeForce RTX 3070Ampere, 8 GB GDDR6 | No, needs 12 GB | Doesn’t fit |
Workstation (4)
| GPU | Fits | Per image |
|---|---|---|
| RTX PRO 6000 Blackwell Workstation EditionBlackwell, 96 GB GDDR7 | Yes | ≈ 0.70–0.72 s |
| RTX PRO 6000 Blackwell Max-Q Workstation EditionBlackwell, 96 GB GDDR7 | Yes | ≈ 0.75–0.77 s |
| RTX 6000 Ada GenerationAda Lovelace, 48 GB GDDR6 | Yes | ≈ 1.1 s |
| RTX A6000Ampere, 48 GB GDDR6 | Yes | ≈ 1.9 s |
Data center (6)
| GPU | Fits | Per image |
|---|---|---|
| B200 (HGX, one GPU)Blackwell, 180 GB HBM3E | Yes | ≈ 0.32–0.35 s |
| H200 SXMHopper, 141 GB HBM3e | Yes | ≈ 0.46–0.49 s |
| H100 SXMHopper, 80 GB HBM3 | Yes | ≈ 0.50–0.53 s |
| A100 SXM 80GBAmpere, 80 GB HBM2e | Yes | ≈ 1.3 s |
| L40SAda Lovelace, 48 GB GDDR6 | Yes | ≈ 1.1 s |
| L4Ada Lovelace, 24 GB GDDR6 | Yes | ≈ 3.2 s |
Where a call's time goes on our RTX 4090: tensor cores, BF16 59%, memory-bound 30%, CUDA cores, FP32 9%, CPU and kernel launches 2%.
GeForce (13)
| GPU | Fits | Per image |
|---|---|---|
| GeForce RTX 5090Blackwell, 32 GB GDDR7 | Yes | ≈ 3.0 s |
| GeForce RTX 5080Blackwell, 16 GB GDDR7 | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 5070 TiBlackwell, 16 GB GDDR7 | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 5070Blackwell, 12 GB GDDR7 | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 5060 Ti 16GBBlackwell, 16 GB GDDR7 | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 5060Blackwell, 8 GB GDDR7 | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 4080Ada Lovelace, 16 GB GDDR6X | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 4070 TiAda Lovelace, 12 GB GDDR6X | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 4070Ada Lovelace, 12 GB GDDR6X | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 3090Ampere, 24 GB GDDR6X | Yes | ≈ 7.9–8.0 s |
| GeForce RTX 3080Ampere, 10 GB GDDR6X | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 3070 TiAmpere, 8 GB GDDR6X | No, needs 18 GB | Doesn’t fit |
| GeForce RTX 3070Ampere, 8 GB GDDR6 | No, needs 18 GB | Doesn’t fit |
Workstation (4)
| GPU | Fits | Per image |
|---|---|---|
| RTX PRO 6000 Blackwell Workstation EditionBlackwell, 96 GB GDDR7 | Yes | ≈ 1.8 s |
| RTX PRO 6000 Blackwell Max-Q Workstation EditionBlackwell, 96 GB GDDR7 | Yes | ≈ 2.0 s |
| RTX 6000 Ada GenerationAda Lovelace, 48 GB GDDR6 | Yes | ≈ 2.8 s |
| RTX A6000Ampere, 48 GB GDDR6 | Yes | ≈ 5.1 s |
Data center (6)
| GPU | Fits | Per image |
|---|---|---|
| B200 (HGX, one GPU)Blackwell, 180 GB HBM3E | Yes | ≈ 0.83–0.86 s |
| H200 SXMHopper, 141 GB HBM3e | Yes | ≈ 1.2 s |
| H100 SXMHopper, 80 GB HBM3 | Yes | ≈ 1.3 s |
| A100 SXM 80GBAmpere, 80 GB HBM2e | Yes | ≈ 3.5 s |
| L40SAda Lovelace, 48 GB GDDR6 | Yes | ≈ 3.0 s |
| L4Ada Lovelace, 24 GB GDDR6 | Yes | ≈ 8.5–8.6 s |
Alternatives
Something wrong on this page, such as a license or a number? Email us.