Adobe Firefly

Generates images from natural language text prompts (up to 750 characters) by routing requests to user-selected generative models—either Adobe's proprietary models or partner models from Google, OpenAI, and Runway. The system enforces client-side prompt length validation and presents a model selection dropdown, but the backend routing logic, latency characteristics, and specific model versions are undisclosed. Output images are returned in standard raster formats for immediate use or refinement in Creative Cloud applications.

Extends or modifies portions of existing images by accepting an image file plus a text prompt describing desired changes, then synthesizing new content that blends seamlessly with the original. The capability integrates directly into Adobe Photoshop's workflow, allowing users to select regions and apply generative fill without creating new layers or destructive edits. Implementation details—such as inpainting architecture, blending algorithms, or how context from the original image is preserved—are undisclosed.

Accepts natural language text prompts (up to 750 characters maximum, enforced client-side) as the primary input method for all generative capabilities (images, video, audio, text effects). The system validates prompt length and rejects inputs exceeding the limit, requiring users to simplify or split complex requests. Prompt engineering guidance, examples, or optimization tools are not mentioned.

Transforms text into stylized visual effects by accepting text input and optional style parameters, then generating rendered text with applied effects (shadows, glows, textures, 3D extrusions, etc.). The capability is mentioned in the product description but not detailed on the website; implementation approach, supported effect types, and integration points are undisclosed. Output is likely a raster image or vector graphic suitable for export to design applications.

Recolors vector graphics by accepting a vector file and color specification (or descriptive color intent), then intelligently remapping colors while preserving vector structure and layer hierarchy. The capability is mentioned in the product description but implementation details are undisclosed; it is unclear whether recoloring is rule-based (e.g., hue-shift), AI-driven (semantic color understanding), or hybrid. Output is a modified vector file in standard formats (SVG, AI, etc.).

Generates video clips from natural language text prompts by routing requests to generative video models (likely Runway or other partner models, as Adobe's own video generation capability is not confirmed). The system accepts text descriptions and returns video files in unspecified formats and durations. Implementation details—such as model selection, video length limits, frame rate, resolution options, and latency—are completely undisclosed.

Generates audio clips and sound effects from natural language text descriptions by routing requests to generative audio models (provider unknown, likely partner models). The system accepts text prompts and returns audio files in unspecified formats and durations. Implementation details—such as audio model selection, duration limits, sample rate, codec, and latency—are completely undisclosed.

Translates video content into target languages while preserving visual elements, likely by detecting and translating audio/subtitles and potentially re-synthesizing speech in the target language. The capability is mentioned for 'content creators' but implementation details—such as supported languages, audio re-synthesis approach, subtitle handling, and quality—are completely undisclosed. Output is a modified video file with translated audio and/or subtitles.

Provides a collaborative workspace (Firefly Boards) for brainstorming and iterating on generative AI outputs, allowing teams to organize generated assets, mix and layer outputs, and iterate on concepts. The boards integrate with Photoshop and Adobe Express for refinement and export. Implementation details—such as collaboration features (real-time vs. async), version control, asset organization, and export formats—are undisclosed.

Offers explicit model provider selection between Adobe's proprietary 'Commercially safe' models (trained on licensed content) and partner models from Google, OpenAI, and Runway, with claimed IP indemnification for Adobe models. The system presents a dropdown selector allowing users to choose between model sources based on their content safety and licensing requirements. Actual indemnification terms, training data sources, and legal coverage are undisclosed.

Integrates Firefly-generated assets directly into Adobe Photoshop and Adobe Express workflows, allowing users to refine, edit, and export generated content without leaving the Creative Cloud ecosystem. Generated images, videos, and audio can be imported as layers, smart objects, or project elements, enabling non-destructive editing and iteration. Integration mechanism—such as native plugins, API calls, or file-based import—is undisclosed.