# US 12,770,305 B2 · One-step diffusion distillation via deep equilibrium models

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,770,305 B2
- Document ID: US12770305B2
- Status: Granted
- Kind: B2
- Application number: 18/332,355
- Filed: 2023-06-09
- Published: 2024-12-12
- Granted: 2026-10-20
- Earliest claimed: 2023-06-09
- Assignees: Robert Bosch GmbH
- Applicants: Robert Bosch GmbH
- Inventors: Zhengyang Geng; Ashwini Polke; Jeremy Kolter; Bahare Azari; Ivan Batalov; Filipe Condessa
- Web page: https://patentgrep.com/patent/US12770305B2

## CPC

- G06T 11/60: Creating or editing images; Combining images with text
- G06T 5/70: Denoising; Smoothing
- G06T 2207/20084: Artificial neural networks [ANN]
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL

## Abstract

Generative equilibrium transformers are disclosed. Disclosed embodiments provide a simple and effective technique that can distill a multi-step diffusion process into a single-step generative model using solely noise/image pairs.

## Claims

### Claim 1

A method comprising:
converting noise into a sequence of two-dimensional patches;
producing a noise embedding vector from the sequence of two-dimensional patches;
tokenizing the noise embedding vector via an injection transformer;
inputting the tokenized noise into an equilibrium transformer;
solving a fixed point via the equilibrium transformer; and
decoding the fixed point to generate an image sample.

### Claim 2 (depends on 1)

The method of claim 1, wherein the injection transformer includes a sequence of injection transformer blocks, the equilibrium transformer includes a sequence of equilibrium transformer blocks differing at an injection interface.

### Claim 3 (depends on 2)

The method of claim 2, wherein each transformer block includes:
a first layer normalization component that receives the noise embedding vector;
a second layer normalization component;
a multi-head attention component between the first layer normalization component and the second layer normalization component; and
a multi-layer perceptron component.

### Claim 4 (depends on 2)

The method of claim 2, wherein the injection interface enables interaction between latent tokens and class embeddings.

### Claim 5 (depends on 4)

The method of claim 4, wherein each transformer block includes:
a first layer normalization component that receives the tokenized noise;
a second layer normalization component;
a multi-head attention component between the first layer normalization component and the second layer normalization component; and
a multi-layer perceptron component.

### Claim 6 (depends on 1)

The method of claim 1, wherein the noise is represented by e∈RH×W×C and the two-dimensional patches are represented by p∈RN×(P2·C), where C is a number of channels, P is the size of patch, H and W denote a height and a width of an original image, and N=HW/P2 is a number of the two-dimensional patches.

### Claim 7 (depends on 1)

The method of claim 1, further comprising:
converting a class label into a class embedding vector;
inputting the class embedding vector into the injection transformer prior to the tokenizing of the noise embedding vector; and
inputting the class embedding vector into the equilibrium transformer prior to the solving of the fixed point.

### Claim 8

A non-transitory memory including computer-executable instructions that when executed by a system cause the system to perform operations including:
converting noise into a sequence of two-dimensional patches;
producing a noise embedding vector from the sequence of two-dimensional patches;
tokenizing the noise embedding vector via an injection transformer;
inputting the tokenized noise into an equilibrium transformer;
solving a fixed point via the equilibrium transformer; and
decoding the fixed point to generate an image sample.

### Claim 9 (depends on 8)

The memory of claim 8, wherein the injection transformer includes a sequence of injection transformer blocks, the equilibrium transformer includes a sequence of equilibrium transformer blocks differing at an injection interface.

### Claim 10 (depends on 9)

The memory of claim 9, wherein each transformer block includes:
a first layer normalization component that receives the noise embedding vector;
a second layer normalization component;
a multi-head attention component between the first layer normalization component and the second layer normalization component; and
a multi-layer perceptron component.

### Claim 11 (depends on 9)

The memory of claim 9, wherein the injection interface enables interaction between latent tokens and class embeddings.

### Claim 12 (depends on 11)

The memory of claim 11, wherein each transformer block includes:
a first layer normalization component that receives the tokenized noise;
a second layer normalization component;
a multi-head attention component between the first layer normalization component and the second layer normalization component; and
a multi-layer perceptron component.

### Claim 13 (depends on 8)

The memory of claim 8, wherein the noise is represented by e∈RH×W×C and the two-dimensional patches are represented by p∈RN×(P2·C), where C is a number of channels, P is the size of patch, H and W denote a height and a width of an original image, and N=HW/P2 is a number of the two-dimensional patches.

### Claim 14 (depends on 8)

The memory of claim 8, wherein the operations further include:
converting a class label into a class embedding vector;
inputting the class embedding vector into the injection transformer prior to the tokenizing of the noise embedding vector; and
inputting the class embedding vector into the equilibrium transformer prior to the solving of the fixed point.

### Claim 15

A system, comprising:
one or more processors; and
a non-transitory memory including computer-executable instructions that when executed by the one or more processors cause the system to perform operations including:
converting noise into a sequence of two-dimensional patches;
producing a noise embedding vector from the sequence of two-dimensional patches;
tokenizing the noise embedding vector via an injection transformer;
inputting the tokenized noise into an equilibrium transformer;
solving a fixed point via the equilibrium transformer; and
decoding the fixed point to generate an image sample.

### Claim 16 (depends on 15)

The system of claim 15, wherein the injection transformer includes a sequence of injection transformer blocks, the equilibrium transformer includes a sequence of equilibrium transformer blocks differing at an injection interface.

### Claim 17 (depends on 16)

The system of claim 16, wherein each transformer block includes:
a first layer normalization component that receives the noise embedding vector;
a second layer normalization component;
a multi-head attention component between the first layer normalization component and the second layer normalization component; and
a multi-layer perceptron component.

### Claim 18 (depends on 16)

The system of claim 16, wherein the injection interface enables interaction between latent tokens and class embeddings.

### Claim 19 (depends on 18)

The system of claim 18, wherein each transformer block includes:
a first layer normalization component that receives the tokenized noise;
a second layer normalization component;
a multi-head attention component between the first layer normalization component and the second layer normalization component; and
a multi-layer perceptron component.

### Claim 20 (depends on 15)

The system of claim 15, wherein the noise is represented by e∈RH×W×C and the two-dimensional patches are represented by p∈RN×(P2·C), where C is a number of channels, P is the size of patch, H and W denote a height and a width of an original image, and N=HW/P2 is a number of the two-dimensional patches.