Adobe Firefly: New Image Generation Models & Redesigned Web App

Adobe Unveils Next-Generation Firefly AI Models and a Revamped Web Application

On Thursday, Adobe announced the latest advancements within its Firefly family of image generation AI models. This includes a new model specifically designed for vector graphics creation, alongside a completely redesigned web application. The new platform consolidates all of Adobe’s AI models, and incorporates options from competing developers.

Enhanced Image Generation Capabilities with Firefly Image Model 4

According to Adobe, the newly released Firefly Image Model 4 represents a significant leap forward in terms of output quality, generation speed, and user control. It offers refined control over stylistic elements, composition, camera perspectives, and zoom levels.

This model is capable of producing images with resolutions reaching up to 2K. A more powerful variant, Image Model 4 Ultra, has also been introduced. This version excels at rendering intricate scenes containing numerous small details and complex structures.

Increased Computational Power and Style Transfer

Alexandru Costin, VP of Generative AI at Adobe, explained that the models were trained utilizing substantially increased computational resources. This enhancement allows for the creation of more detailed and visually rich images.

Compared to previous iterations, these new models demonstrate improved text rendering within images. Furthermore, they incorporate features enabling users to leverage their own images as stylistic references for generating new visuals.

Firefly Video Model Now Widely Accessible

The Firefly video model, initially released in limited beta last year, is now available to all users. This tool empowers users to generate video clips directly from text prompts or existing images.

Users can specify camera angles, define start and end frames for precise shot control, add atmospheric effects, and customize motion design elements. The model supports video generation at resolutions up to 1080p.

Introducing the Firefly Vector Model

The Firefly Vector Model introduces the ability to create scalable, editable vector artwork. This facilitates the iterative design and generation of variations for logos, product packaging, icons, scenes, and patterns.

A Unified Platform with Access to Multiple AI Models

Adobe’s redesigned web application provides a central hub for accessing all of these models. It also integrates image and video generation models from other leading AI companies, including OpenAI (GPT image generation), Google (Imagen 3 and Veo 2), and Flux (Flux 1.1 Pro).

Users can seamlessly switch between these different models as needed. All generated images will be embedded with content credentials to ensure transparency and traceability. Adobe indicated that further AI models may be added to the platform in the future.

Firefly Boards: A Collaborative Ideation Space

Adobe is currently publicly testing Firefly Boards, a digital canvas designed for brainstorming and moodboarding. This tool allows users to generate or import images, remix them, and collaborate with colleagues – offering functionality similar to AI-powered ideation platforms like Visual Electric, Cove, and Kosmik.

Access to Boards is available through the Firefly web application.

Integration into Adobe’s Product Suite and API Availability

The company has announced plans to integrate these new models into its existing product portfolio. However, a specific timeline for this rollout has not yet been disclosed.

Furthermore, Adobe is making its Text-to-Image API and Avatar API generally available. A new Text-to-Video API is also available for beta testing. These APIs are accessible through the Firefly Services collection.

Content Authenticity and AI Training Opt-Out

Adobe is also piloting Adobe Content Authenticity, a web application that enables users to attach credentials to their creations. This metadata establishes ownership and attribution. Users will also have the option to specify whether their images can be used for training AI models.