Skip to content
ModelsThatRun

SDXL-Lightning

ByteDance's distillation of Stable Diffusion XL to a few steps. The 4-step UNet runs in SDXL base 1.0's pipeline and makes 1024x1024 images.

Size

Conditional: code Apache-2.0, weights LicenseRef-CreativeML-OpenRAIL-PlusPlus-M. Measured on our RTX 4090: 0.56 s per 1024×1024 image at 4 steps, 14.4 GiB VRAM peak. We scored it 0.54 on GenEval. safetensors weights, 6.46 GB.

License

Conditional

Code
Apache-2.0checked 11 Nov 2026
Weights
LicenseRef-CreativeML-OpenRAIL-PlusPlus-Mall sizes, checked 11 Nov 2026

CreativeML Open RAIL++-M, for this UNet and for the SDXL base 1.0 text encoders and VAE it runs with: commercial use is allowed, under the use restrictions in its Attachment A.

On our RTX 4090, measured

0.56s per image

p90 0.56 s, fp16, batch 1, 1024x1024, 4 steps, guidance 0

VRAM peak
14.4 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. EulerDiscreteScheduler, 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

0.54GenEvalmeasured

GenEval, 2,212 images, fp16, GenEval af4902f · 11 Nov 2026

Weights

safetensors

Size
6.46 GB
sha256
39a90a8a2a25c1372633c1e8eb1a96f85eab937bcb58cfc6ec9baf66242ec23c
Download 4step weights

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, 10 marked as failures.

  1. SDXL-Lightning 4step's image for the prompt: Exactly five green apples in a single row on a wooden table, plain white background, product photo. Marked as a failure.

    Counting

    Fails: About a dozen apples scattered over the table, not five in a row
    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.

  2. SDXL-Lightning 4step's image for the prompt: A blue cube resting on top of a red sphere, with a yellow cone standing beside them, on a grey studio floor. Marked as a failure.

    Color binding

    Fails: A red cube alone, with two small red and yellow shapes; no blue cube, sphere or cone
    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.

  3. SDXL-Lightning 4step's image for 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. Marked as a failure.

    Spatial relations

    Fails: Two dogs and a cat, with the cat on a footstool at the right
    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.

  4. SDXL-Lightning 4step's image for the prompt: A wooden sign above a bakery door that reads "FRESH BREAD DAILY" in white painted capital letters. Marked as a failure.

    Text rendering

    Fails: Garbled letters instead of FRESH BREAD DAILY
    The 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.

  5. SDXL-Lightning 4step's image for the prompt: A café chalkboard menu with three lines written in chalk: "Coffee 3", "Tea 2" and "Cake 4". Marked as a failure.

    Text rendering, several lines

    Fails: A full menu of garbled text instead of the three lines
    The 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.

  6. SDXL-Lightning 4step's image for the prompt: Aerial photograph of a winding river through an autumn forest at sunrise, thin mist lying over the water.

    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.

  7. SDXL-Lightning 4step's image for the 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. Marked as a failure.

    Photorealism, food

    Fails: Many slices of pork, and a second halved egg in a side dish
    The 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.

  8. SDXL-Lightning 4step's image for the 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. Marked as a failure.

    Transparency and reflection

    Fails: No street shows through the water in the glass
    The 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.

  9. SDXL-Lightning 4step's image for the prompt: Close-up of an adult's two hands kneading bread dough on a floured wooden counter, no face in the frame.

    Hands

    The 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.

  10. SDXL-Lightning 4step's image for the prompt: Macro photograph of a plain mechanical wristwatch with no brand name, its hands showing ten past ten, the second hand pointing at six. Marked as a failure.

    Fine detail

    Fails: Shows the movement, with no dial or hands
    The 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.

  11. SDXL-Lightning 4step's image for the 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. Marked as a failure.

    Long, dense prompt

    Fails: The lamp is on the left, the books are piled on a table, and there is no teapot, cup or rain
    The 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.

  12. SDXL-Lightning 4step's image for the prompt: A watercolor painting of a lighthouse on a rocky coast under a stormy sky, loose brush strokes, visible paper texture.

    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.

  13. SDXL-Lightning 4step's image for the prompt: An empty park bench in fresh snow, with no people and no birds anywhere.

    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.

  14. SDXL-Lightning 4step's image for the prompt: A giraffe wearing a long knitted red scarf, standing in a snowy city square at night, street lamps glowing. Marked as a failure.

    Unusual combination

    Fails: A red knitted sweater instead of a scarf
    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 SDXL-sized 1024x1024 images in 4 steps instead of 50. Plain scenes held up in our prompts: an aerial landscape, hands, a watercolor and an empty bench.

Avoid it when

You need legible text, exact counts, each color on the right object, or left and right kept apart: it missed all of them in our prompts, as SDXL base did.

Sizes

SizePer imageVRAMAccuracyWeightsRunsLicense
4step0.56 s1024x1024 · 4 steps14.4 GiB0.54 GenEval, measured6.46 GBAs isConditional

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, FP16 (FP32 accumulate) 38%, memory-bound 23%, tensor cores, TF32 21%, CPU and kernel launches 7%, unclassified kernels 7%, CUDA cores, FP32 5%.

GeForce (13)
GPUFitsPer image
GeForce RTX 5090Blackwell, 32 GB GDDR7Yes≈ 0.41–0.44 s
GeForce RTX 5080Blackwell, 16 GB GDDR7Yes≈ 0.70–0.74 s
GeForce RTX 5070 TiBlackwell, 16 GB GDDR7Yes≈ 0.83–0.91 s
GeForce RTX 5070Blackwell, 12 GB GDDR7No, needs 14 GBDoesn’t fit
GeForce RTX 5060 Ti 16GBBlackwell, 16 GB GDDR7Yes≈ 1.5–1.7 s
GeForce RTX 5060Blackwell, 8 GB GDDR7No, needs 14 GBDoesn’t fit
GeForce RTX 4080Ada Lovelace, 16 GB GDDR6XYes≈ 0.82–0.88 s
GeForce RTX 4070 TiAda Lovelace, 12 GB GDDR6XNo, needs 14 GBDoesn’t fit
GeForce RTX 4070Ada Lovelace, 12 GB GDDR6XNo, needs 14 GBDoesn’t fit
GeForce RTX 3090Ampere, 24 GB GDDR6XYes≈ 0.95–1.1 s
GeForce RTX 3080Ampere, 10 GB GDDR6XNo, needs 14 GBDoesn’t fit
GeForce RTX 3070 TiAmpere, 8 GB GDDR6XNo, needs 14 GBDoesn’t fit
GeForce RTX 3070Ampere, 8 GB GDDR6No, needs 14 GBDoesn’t fit
Workstation (4)
GPUFitsPer image
RTX PRO 6000 Blackwell Workstation EditionBlackwell, 96 GB GDDR7Yes≈ 0.25–0.31 s
RTX PRO 6000 Blackwell Max-Q Workstation EditionBlackwell, 96 GB GDDR7Yes≈ 0.27–0.33 s
RTX 6000 Ada GenerationAda Lovelace, 48 GB GDDR6Yes≈ 0.36–0.42 s
RTX A6000Ampere, 48 GB GDDR6Yes≈ 0.64–0.69 s
Data center (6)
GPUFitsPer image
B200 (HGX, one GPU)Blackwell, 180 GB HBM3EYes≈ 0.11–0.22 s
H200 SXMHopper, 141 GB HBM3eYes≈ 0.16–0.26 s
H100 SXMHopper, 80 GB HBM3Yes≈ 0.17–0.27 s
A100 SXM 80GBAmpere, 80 GB HBM2eYes≈ 0.40–0.58 s
L40SAda Lovelace, 48 GB GDDR6Yes≈ 0.38–0.44 s
L4Ada Lovelace, 24 GB GDDR6Yes≈ 0.99–1.1 s

From the authors

Where to read their own account of it.

Alternatives

Something wrong on this page, such as a license or a number? Email us.

⇆
InputOutput