DALLE-2 has a secret language.
"Apoploe vesrreaitais" means birds.
"Contarra ccetnxniams luryca tanniounons" means bugs or pests.
The prompt: "Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons" gives images of birds eating bugs.
A thread (1/n)🧵
A known limitation of DALLE-2 is that it struggles with text. For example, the prompt: "Two farmers talking about vegetables, with subtitles" gives an image that appears to have gibberish text on it.
However, the text is not as random as it initially appears... (2/n)
We feed the text "Vicootes" from the previous image to DALLE-2. Surprisingly, we get (dishes with) vegetables! We then feed the words: "Apoploe vesrreaitars" and we get birds. It seems that the farmers are talking about birds, messing with their vegetables! (3/n)
Another example: "Two whales talking about food, with subtitles". We get an image with the text "Wa ch zod rea" written on it. Apparently, the whales are actually talking about their food in the DALLE-2 language. (4/n)
Some words from the DALLE-2 language can be learned and used to create absurd prompts. For example, "painting of Apoploe vesrreaitais" gives a painting of a bird. "Apoploe vesrreaitais" means to the model "something that flies" and can be used across diverse styles. (5/n)
The discovery of the DALLE-2 language creates many interesting security and interpretability challenges.
Currently, NLP systems filter text prompts that violate the policy rules. Gibberish prompts may be used to bypass these filters. (6/n)
Based on valid comments, we updated our paper with a discussion on Limitations and changed the title to Discovering the Hidden Vocabulary of DALLE-2. Thanks to @mraginsky@rctatman@benjamin_hilton and others for useful comments.
Given a few images, we learn pseudo-words that represent a concept at different resolutions.
"A painting of a dog in the style of <jane(number)>" gives different levels of artistic freedom to match the <jane> style based on the number index.
The key idea of our method is to condition the embedding of the learned concept on the diffusion time.
Instead of learning one embedding to represent the concept, we learn a set of embeddings: each element of the set represents the object at different resolutions.
During inference, we can use the embeddings in many creative ways to access the learned object at different resolutions.
For example, given a painting made of buttons, we can isolate the buttons and create new objects with that texture.
Announcing Soft Diffusion: A framework to correctly schedule, learn and sample from general diffusion processes.
State-of-the-art results on CelebA, outperforms DDPMs and vanilla score-based models.
A 🧵to learn about Soft Score Matching, Momentum Sampling and the role of noise
Typically, diffusion models generate images by reversing a known corruption process that gradually adds noise.
We show how to learn to reverse diffusions that involve a linear deterministic degradation and a stochastic part (additive noise).
Ingredient 1: Soft Score Matching.
Soft Score Matching incorporates the filtering process in the network. It trains the model to predict an image that after corruption matches the diffused observation.