The B-roll (auxiliary screen) in short videos is used to enrich the visual hierarchy and help the audience understand the oral broadcast. However, the traditional production process is trivial and requires multiple processes such as material search, composition arrangement, animation and editing. Open source projects on GitHub gbro-collage-broll It provides a lightweight path: input a piece of spoken word copy for about 5 seconds, directly generate a half-tone paper collage animation with a magazine editing style, and leave the tedious links from concept conception to camera movement generation to AI for processing.
Visually, the project focuses on a retro editorial style: using highly saturated solid colors as a base, paired with black and white half-tone photo cutouts and colorful cardboard elements, which is close to the hand-collage texture of an art magazine. The animation does not use the common linear fade-in or slow push-and-pull, but starts from a blank screen. Each layer of clipping elements slides in, positions and snaps into the composition in sequence, with a frame-jumping feel like stop-motion animation. This form is very suitable for oral broadcast scenarios with many abstract concepts and lack of real-life scenes (such as technological interpretation, business analysis, and knowledge-based oral broadcasts). It can make the content that can only rely on subtitles fall into concrete visual metaphors.
In order to prevent AI's random ideas from deviating from the textual meaning, gbro-collage-broll introduced a "three-gate approval" mechanism to block artificial aesthetics at key nodes:
- metaphorical confirmation: The system first refines the core concept, main objects, background color and screen assembly sequence based on the copywriting. At this stage, only the text plan will be produced, and no graphics and view generation quota will be consumed.
- Still frame confirmation: After the plan is approved, the system generates a collage still frame and a multi-view contact list, and the creator confirms the composition and texture.
- video generation: Call the first and last frame control capabilities of Gemini Omni Flash to render a 5-second assembly animation, and automatically perform verification such as blanking the first frame, extracting frames second by second, and aligning the last frame.
This segmented release mechanism leaves creative adjustments at the text and picture stage, with extremely low trial and error costs, and prevents biased designs from directly rushing into high-cost video rendering.
The project is packaged in the form of Agent Skill and runs in an AI Agent environment that supports the skill mechanism. The running environment requires Python 3.10+, Gemini API Key, google-genai library, and audio and video tools such as ffmpeg/ffprobe; the still frame generation uses the image_gen tool of the Codex environment to dynamically generate the Gemini API. The output format is fixed to standard mobile phone vertical screen specifications: 9:16, 5 seconds duration, 720×1280 resolution, 24fps silent MP4 video, which can be directly dragged into the editing track as a picture-in-picture or transition cutter. During installation, you only need to clone the warehouse to the Agent skill directory, enter "collage b-roll" or "paper collage b-roll" and add copywriting in the dialog to evoke the process.
The core orientation of this workflow is not to be "fully automatic and hands-off", but to divide the rights and responsibilities of humans and machines: humans will check whether the metaphor is appropriate, whether the theme matches, and whether the composition is eye-catching; mechanical work such as material cutout, layered alignment, and motion effect frame insertion are left to the model. In the face of complex abstract narratives, a hand-made collage that fits the context is often more eye-catching than a bunch of generic video materials with no narrative connection.
gbro-collage-broll avoids the harsh realism often found in current AI video tools, and uses the blank space and cutting texture of collage art to resolve the unnaturalness of the generated images. For short video authors, it can not only save a lot of time in searching for B-roll in the material library, but also ensure that each animation accurately fits their oral intention.
Project address:https://github.com/pyang5166/gbro-collage-broll