GitHub Copilot Translates Mockups Into Code with Image-to-Code Generation
GitHub has unveiled a series of updates for Copilot, offering a preview of a more autonomous future for its AI-powered programming assistant.
A key new feature is Copilot Vision, which lets users upload a screenshot, photo, or diagram directly into the chat. Copilot then uses the visual input to generate corresponding interface elements, code, or descriptive alt text.
For instance, a marketing team member could capture a webpage screenshot and visually indicate desired changes. Instead of describing the edits in text, they can simply upload the annotated image and instruct Copilot to implement the shown modifications.
It's important to note that this functionality was previously available as a VS Code extension starting last October. Microsoft had indicated it would eventually be deprecated in favor of a native feature within GitHub Copilot Chat. The earlier version also required users to supply their own API key.

Vision for Copilot | Image Credits: GitHub GitHub is also launching "next edit suggestions," an enhancement to Copilot's core code completion capabilities.
Previously, GitHub Copilot in the VS Code editor primarily worked from the cursor's current position. Now, it analyzes recent edits across the file to predict the developer's likely next action, using richer contextual signals to automatically propose the subsequent edit.
Developers can accept these suggestions by pressing the Tab key or reject them by pressing Esc.

GitHub Copilot Next Edit Suggestions | Image Credits: GitHub GitHub's Autonomous Future
GitHub first introduced Copilot Edits last November, providing a natural language interface for executing multi-file changes. This feature is now generally available and includes a new "agent mode." This mode automatically identifies all files relevant to the developer's requested changes, eliminating the need to manually specify which files should be modified.
"Copilot now does more work to understand the intent behind your original request and then tries to solve it," GitHub CEO Thomas Dohmke explained to TechCrunch.
Long-term, the vision is to apply similar "agent modes" to other areas of Copilot, enabling greater automation across interconnected development tasks. This direction is underscored by a new initiative called Project Padawan, essentially a software engineering (SWE) agent capable of independently handling entire tasks when a developer assigns an issue to Copilot.
Dohmke did not provide a specific timeline for Project Padawan's full release, stating only that GitHub will collaborate with partners and the community to refine it.
"We're seeking partners who are also building agents to integrate into that workflow, and customers to work with us and provide feedback—because we know it won't be perfect at this stage," Dohmke said.
TechCrunch publishes a dedicated AI newsletter. Sign up here to receive it in your inbox every Wednesday.
Related article
Microsoft retires failed AI tools and consolidates Copilot apps
Two years ago, Microsoft characterized AI as a “generational shift” in technology that it aimed to lead. Today, the company is consolidating its Copilot-branded consumer and business applications, while discontinuing several underperforming AI featur
Haleon and Microsoft Partner to Automate Global Health
Haleon is leveraging Microsoft AI to accelerate scientific discoveries in everyday health, including dental care. (Credit: Haleon)The UK-based consumer health firm Haleon has entered a partnership with Microsoft, a step that Claire Dickson, its Chief
Microsoft pulls back Copilot AI bloat on Windows
On Friday, Microsoft announced a series of updates aimed at enhancing the quality of its Windows 11 operating system, most notably reducing the number of access points for its AI assistant, Copilot.The company stated that it will scale back Copilot A
Related Special Topic Recommendations
Comments (2)
0/500
Ah, l'idée de transformer une simple capture d'écran en code via Copilot, c'est séduisant sur le papier, mais franchement, je me demande si ça va vraiment marcher sans générer un tas de bugs invisibles. On dirait un gadget de plus pour faire joli dans les démos, mais dans la vraie vie, le diable se cache dans les détails. Peur que ça finisse par produire du code spaghetti que personne ne voudra relire. 😅
GitHub has unveiled a series of updates for Copilot, offering a preview of a more autonomous future for its AI-powered programming assistant.
A key new feature is Copilot Vision, which lets users upload a screenshot, photo, or diagram directly into the chat. Copilot then uses the visual input to generate corresponding interface elements, code, or descriptive alt text.
For instance, a marketing team member could capture a webpage screenshot and visually indicate desired changes. Instead of describing the edits in text, they can simply upload the annotated image and instruct Copilot to implement the shown modifications.
It's important to note that this functionality was previously available as a VS Code extension starting last October. Microsoft had indicated it would eventually be deprecated in favor of a native feature within GitHub Copilot Chat. The earlier version also required users to supply their own API key.

GitHub is also launching "next edit suggestions," an enhancement to Copilot's core code completion capabilities.
Previously, GitHub Copilot in the VS Code editor primarily worked from the cursor's current position. Now, it analyzes recent edits across the file to predict the developer's likely next action, using richer contextual signals to automatically propose the subsequent edit.
Developers can accept these suggestions by pressing the Tab key or reject them by pressing Esc.

GitHub's Autonomous Future
GitHub first introduced Copilot Edits last November, providing a natural language interface for executing multi-file changes. This feature is now generally available and includes a new "agent mode." This mode automatically identifies all files relevant to the developer's requested changes, eliminating the need to manually specify which files should be modified.
"Copilot now does more work to understand the intent behind your original request and then tries to solve it," GitHub CEO Thomas Dohmke explained to TechCrunch.
Long-term, the vision is to apply similar "agent modes" to other areas of Copilot, enabling greater automation across interconnected development tasks. This direction is underscored by a new initiative called Project Padawan, essentially a software engineering (SWE) agent capable of independently handling entire tasks when a developer assigns an issue to Copilot.
Dohmke did not provide a specific timeline for Project Padawan's full release, stating only that GitHub will collaborate with partners and the community to refine it.
"We're seeking partners who are also building agents to integrate into that workflow, and customers to work with us and provide feedback—because we know it won't be perfect at this stage," Dohmke said.
TechCrunch publishes a dedicated AI newsletter. Sign up here to receive it in your inbox every Wednesday.
Microsoft retires failed AI tools and consolidates Copilot apps
Two years ago, Microsoft characterized AI as a “generational shift” in technology that it aimed to lead. Today, the company is consolidating its Copilot-branded consumer and business applications, while discontinuing several underperforming AI featur
Haleon and Microsoft Partner to Automate Global Health
Haleon is leveraging Microsoft AI to accelerate scientific discoveries in everyday health, including dental care. (Credit: Haleon)The UK-based consumer health firm Haleon has entered a partnership with Microsoft, a step that Claire Dickson, its Chief
Microsoft pulls back Copilot AI bloat on Windows
On Friday, Microsoft announced a series of updates aimed at enhancing the quality of its Windows 11 operating system, most notably reducing the number of access points for its AI assistant, Copilot.The company stated that it will scale back Copilot A
Ah, l'idée de transformer une simple capture d'écran en code via Copilot, c'est séduisant sur le papier, mais franchement, je me demande si ça va vraiment marcher sans générer un tas de bugs invisibles. On dirait un gadget de plus pour faire joli dans les démos, mais dans la vraie vie, le diable se cache dans les détails. Peur que ça finisse par produire du code spaghetti que personne ne voudra relire. 😅





Home






