Old Models Foundation

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

A visual generator and a model that connects images to language form a feedback loop. The image is repeatedly nudged toward a prompt, producing the intricate, painterly scenes of an early text-to-image era.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.

Grows a texture from a sample image by matching pixel neighborhoods. Published in September 1999. This is a non-neural synthesis algorithm.

Synthesizes textures by joining overlapping patches from a source image. The 2001 method also supports texture transfer. It does not use pretrained neural weights.

Variational autoencoders were introduced in December 2013. This sandbox uses Kingma’s author-published 2014 MNIST weights in a NumPy reconstruction, producing 28 × 28 handwritten digits without retraining.

Introduced in February 2015, DRAW generates images through successive canvas updates. This sandbox uses Eric Jang’s pretrained 2016 MNIST implementation in a NumPy reconstruction, distinct from the original DeepMind weights. This checkpoint produces mostly loop-shaped digits.

DeepDream amplifies the patterns a neural network recognizes inside an image. Choose an ImageNet object network or a Places205 or Places365 scene network, then adjust the patterns and detail scales it amplifies. The original 2015 DeepDream used GoogLeNet; the other networks apply the same technique to different learned features.

Introduced in November 2015, DCGAN showed how convolutional adversarial networks could learn visual structure from photographs. Explore the original Torch release’s LSUN bedroom weights, recovered in a modern runtime, generating tiny 64 × 64 rooms from random seeds.

Introduced in a 2016 paper, pix2pix learns to translate between paired images. Its demonstrations turn edges into shoes or handbags and building labels into facades, making a drawing the starting point for generation.

BigGAN was introduced in September 2018. This sandbox uses the pretrained BigGAN-deep 512px ImageNet generator, conditioned on one of 1,000 image classes. Truncation controls variation in the sampled input.

NVIDIA’s original StyleGAN, introduced in a December 2018 paper, creates images from random inputs and controls visual features at different scales. Explore original checkpoints for faces, cats, bedrooms and cars, mixing their styles and changing fine details.

An independent open project that began in 2021, building images from sequences of visual tokens. This exhibit preserves Mini itself, distinct from later Mega checkpoints, current Craiyon and OpenAI’s DALL·E.

A community notebook that combines diffusion models with CLIP guidance to turn text prompts into images. The project began in October 2021; its diffusion network, guidance models and image cutouts together shape the result.

Available by April 2022, this CompVis predecessor builds images in a compressed space before decoding them into pixels. Its BERT text conditioning makes it a distinct system, rather than another Stable Diffusion checkpoint.

The larger open model from the team behind DALL·E Mini and Craiyon, released in 2022. This exhibit uses the original public Mega checkpoint and VQGAN decoder, with one sample and no CLIP reranking. It is distinct from today’s proprietary Craiyon models.

The recommended checkpoint at Stable Diffusion’s August 2022 public launch. Its release helped turn text-to-image generation into something people could run, modify and build upon themselves.

ControlNet uses an image to guide a diffusion model. Canny follows its outlines; QR Monster blends light and dark patterns into a scene, including spiral and checkerboard illusions. Both modes use Stable Diffusion 1.5.