擒数网 （随信APP) | 挑战Sora！Adobe发布人工智能视频工具，将视频处理变得轻而易举。

Recently, the AI video field has been bustling, and Adobe has just joined this melee.

At this year's Adobe Max conference, Adobe officially launched its own AI video model - Firefly Video Model, entering the realm of generative artificial intelligence.

Adobe's Chief Product Marketing Manager Meagan Keane stated that the launch of Firefly Video Model aims to simplify and accelerate video creation, as well as enhance the storytelling capabilities of videos. In summary, this model includes three interesting and practical features:

• Generative Extend
• Text-to-Video
• Image-to-Video

Among them, Generative Extend has already been integrated into Adobe's professional video editing software Premiere Pro; Text-to-Video and Image-to-Video have just been released in beta version on the Firefly Web app.

As the name suggests, if your video clips are too short or missing certain parts, Generative Extend can help you "generatively extend" the beginning or end of the video clip.

In other words, if the required footage is not available, or if a clip's editing timing is too early or too late, the editor only needs to drag the beginning or end of the video clip, and Generative Extend can automatically fill in generative content to maintain the video's continuity and smoothness.

This means that if a video creator obtains a "useless shot" during the shooting process, they may no longer need to return to the original location for reshooting, but can use Generative Extend to directly extend and repair the original clip.

To a certain extent, this can indeed help improve video creation efficiency - of course, provided that the quality of the segments extended by Generative Extend is solid.

Unfortunately, the current maximum resolution limit of this function is 1080p, and the maximum expansion of a segment is only two seconds.

Therefore, Generative Extend is only suitable for making minor adjustments to video clips, meaning it can help users fix certain details, but cannot replace the creators themselves for extensive creation.

Generative Extend can also help users make adjustments during the shooting process, correcting eye shifts or other unexpected movements that occur during shooting.

Aside from videos, Generative Extend can also be used for smooth editing of audio. It can extend environmental background sound effects in a video for up to ten seconds, but cannot extend dialogues or music.

If you want to save some effort in the video creation process, you can directly generate videos using the Text-to-Video feature. Like its "predecessors" Runway and OpenAI's Sora, users only need to input text descriptions of the videos they want to generate, and the function can simulate various styles such as "real film," "3D animation," and "stop-motion animation" to create corresponding video clips.

Here are some video clips generated using the Text-to-Video feature for you to experience:

In addition to realistic, cinematic visual images, Text-to-Video also supports generating some "abstract" scenes. For example, it can be used to generate videos with elements like flames, water, lens flares, and smoke, which can be overlaid on existing videos to enhance the visual depth and interest of the content.

It is worth mentioning that the video clips generated using Text-to-Video can also be further refined using a series of "camera controls" that can simulate real camera angles, movement, and shooting distances.

The Image-to-Video function goes even further, allowing users to add "reference images" in the video creation process to generate videos that are closer to what the user imagines.

Video creators can start from a single image or photo and use the Image-to-Video function to directly produce "B-roll" (a filmmaking term for supplementary video footage other than the main footage) videos.

Users can also upload a single frame of a video and let Image-to-Video automatically fill in missing frames to create complete close-up shots.

"It is well known that videos cannot be Photoshopped, so they must be real."

However, the "cyber reshoot" function of Image-to-Video makes this statement a thing of the past. It acts like a "God's hand" that allows users to directly manipulate and modify video content (such as character actions in the video). Take a look at the examples below:

Currently, the maximum length of the video clips generated by the Text-to-Video and Image-to-Video functions is only five seconds, with the highest quality being 720p and 24 frames per second. The video generation time is approximately 90 seconds, and Adobe is developing a "turbo mode" to shorten this time.

The insufficient quality of the generated videos indicates that users still cannot use the Firefly Video Model to generate a complete film, and can only use it as an auxiliary tool for creation.

Adobe emphasizes that the Firefly Video Model is only trained on licensed content (such as Adobe Stock) and public domain content, and will not be used on customer-provided content.

In addition, videos created or edited using the Firefly Video Model can embed content credentials, intended to provide attribution proof for creators and content sources, helping to declare AI usage and ownership rights to ensure "business security."

Interested readers can join the Adobe Firefly Video Model experience waitlist at the following link:
https://www.adobe.com/products/firefly/features/ai-video-generator.html