# US 2025/0147648 A1 · Systems and methods for generating and using training data > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2025/0147648 A1 - Document ID: US20250147648A1 - Status: Published application - Kind: A1 - Application number: 18/931,721 - Filed: 2024-10-30 - Published: 2025-05-08 - Earliest claimed: 2023-11-08 - Assignees: Canva Pty Ltd - Applicants: Canva Pty Ltd - Inventors: Haitao Li - Web page: https://patentgrep.com/patent/US20250147648A1 ## CPC - G06F 3/04845: for image manipulation, e.g. dragging, rotation, expansion or change of colour - G06T 11/00: Two-dimensional [2D] image generation - G06F 3/0483: Interaction with page-structured environments, e.g. book metaphor - G06T 2207/20081: Training; Learning - G06T 2207/20084: Artificial neural networks [ANN] - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) - G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL ## Abstract Computer-implemented methods for forming a training dataset for a machine learning model are described. The methods include receiving an intent object and two or more design snapshots associated with the intent object, where the intent object indicates an intended design outcome and the two or more design snapshots include at least a first design snapshot indicating an initial state of a design and a second design snapshot indicating a state of the design after the intended design outcome is achieved. One or more design edits made to the design based on the first design snapshot and the second design snapshot are identified. And a training datapoint based on the intent object and the identified one or more design edits is generated. ## Claims ### Claim 1 A computer-implemented method, comprising: forming a training dataset for a machine learning model by: a) receiving an intent object and two or more design snapshots associated with the intent object, wherein the intent object indicates an intended design outcome and the two or more design snapshots include at least a first design snapshot indicating an initial state of a design and a second design snapshot indicating a state of the design after the intended design outcome is achieved; b) identifying one or more design edits made to the design based on the first design snapshot and the second design snapshot; and c) generating a training datapoint based on the intent object and the identified one or more design edits. ### Claim 2 (depends on 1) The computer-implemented method of claim 1, further comprising repeating steps a)-c) multiple times for the same design to generate a sequence of the training datapoints; wherein each datapoint in the sequence is associated with the same design. ### Claim 3 (depends on 2) The computer-implemented method of claim 2, further comprising generating multiple sequences of the training datapoints for different designs. ### Claim 4 (depends on 1) The computer-implemented method of claim 1, further comprising repeating steps a)-c) multiple times for the same intent object. ### Claim 5 (depends on 1) The computer-implemented method of claim 1, wherein the training dataset is used for training an ML model to output one or more design edits based on an input intent object. ### Claim 6 A computer-implemented method for forming a training dataset for a machine learning model, the method comprising: receiving a first intent object associated with a first design snapshot and a second design snapshot; receiving a second intent object associated with the second design snapshot and a third design snapshot; determining a third intent object based on a combination of the first and second intent objects; and associating with the third intent object the first design snapshot and the third design snapshot, wherein each of the first, second and third intent objects and their associated snapshots are included in the training dataset. ### Claim 7 (depends on 6) The computer-implemented method of claim 6, further comprising determining design edits associated with each of the first, second and third intent objects, wherein determining the design edits for each intent object is based on the design snapshots associated with that intent object. ### Claim 8 (depends on 7) The computer-implemented method of claim 7, wherein the training dataset is used for training a ML model to output one or more design edits based on an input intent object. ### Claim 9 (depends on 6) The computer-implemented method of claim 6, wherein the third intent object is determined by an ML model. ### Claim 10 (depends on 9) The computer-implemented method of claim 9, wherein the third intent object is a higher level intent object than at least one of the first intent object and the second intent object. ### Claim 11 A computer-implemented method, comprising: forming a training dataset for a machine learning model by: a) receiving an intent object and two or more design snapshots associated with the intent object, wherein the intent object indicates an intended design outcome and the two or more design snapshots include at least a first design snapshot indicating an initial state of a design and a second design snapshot indicating a state of the design after the intended design outcome is achieved; and b) generating a training datapoint including the intent object and the two or more design snapshots. ### Claim 12 (depends on 11) The computer-implemented method of claim 11, further comprising repeating steps a) and b) multiple times for a same design to generate a sequence of training datapoints; wherein each training datapoint in the sequence is associated with the same design. ### Claim 13 (depends on 12) The computer-implemented method of claim 12, further comprising generating multiple sequences of the training datapoints for different designs. ### Claim 14 (depends on 11) The computer-implemented method of claim 11, wherein the training dataset is used for training an ML model to output one or more design commands based on an input intent object. ### Claim 15 (depends on 12) The computer-implemented method of claim 12, further comprising: retrieving a first training datapoint and a second training datapoint from the sequence of training datapoints, the first training datapoint including a first intent object and a first and second design snapshots and the second training datapoint including a second intent object and third and fourth design snapshots; combining the first intent object with the second intent object to generate a third intent object; associating with the third intent object the first design snapshot and the fourth design snapshot; and generating an additional training datapoint to include in the sequence of training datapoints, the additional training datapoint including the third intent object and the first design snapshot and fourth design snapshot. ### Claim 16 (depends on 15) The computer-implemented method of claim 15, wherein the third intent object is determined by an ML model. ### Claim 17 (depends on 15) The computer-implemented method of claim 15, wherein the third intent object is a higher level intent object than at least one of the first intent object and the second intent object. ### Claim 18 (depends on 11) The computer-implemented method of claim 11, further comprising determining design edits associated with each intent object, wherein determining the design edits for each intent object is based on the design snapshots associated with that intent object. ### Claim 19 (depends on 11) The computer-implemented method of claim 11, wherein the training datapoints are used for training an ML model to output one or more design edits based on an input intent object.