Summary
Highlights
The speaker introduces the comprehensive process of creating an AI advertisement for Osulloc, emphasizing that a detailed paper outlining the steps will be shared with viewers. The goal is to explain how to create a story-driven video using AI, focusing on the overall workflow rather than just tool usage.
The project took approximately 16 days, with 8 days dedicated to planning (4 for concept, 4 for storyboard) and 8 days for production (5 for image generation, 3 for video editing). This highlights that planning and conceptualization are as crucial as the actual AI generation process.
The concept planning stage is about clearly defining the video's purpose and spending ample time on research. The speaker advises against limiting creativity by preconceived notions of AI capabilities or budget. The strategy was to imagine diverse characters and situations (six target audiences for six products) and incorporate subtle AI elements to create a 'real yet fantastical' tone, using split cuts and AI-generated backgrounds to enhance the narrative.
Following concept planning, a storyboard is created to visualize the video's flow, transitions, and narrative development. This stage involves extensive research and analysis of reference videos and images to craft specific scenes, camera angles, and character interactions without pre-determining AI feasibility. This ensures the story drives the visuals, not the other way around.
The next step is generating 'key cuts' using Midjourney, which provides the highest quality images. These main cuts serve as the central visual references for each scene, ensuring consistency in character appearance, props, and overall atmosphere throughout the video. Examples of key cuts for different characters and settings are shown, emphasizing their importance in guiding subsequent image generation.
Most of the detailed images are generated using Nanovanna Pro, leveraging the key cuts and storyboards. This involves using existing images, applying the generated key cuts to change characters or scenes, and refining details through inpainting. The emphasis is on iteratively generating and refining images to match the desired vision, often creating multiple versions to select the best one.
For voice acting, live voice actors were hired via platforms like KMONG to ensure a natural and immersive experience, avoiding the artificiality of AI voices. For music, Artlist was used for its extensive library, although the video suggests AI music generation tools like Suno could also be a viable option, especially with reference tracks.
An animatic (pre-animation) is created using static images, voiceovers, and music to check the video's timing and flow before full animation. This helps identify and correct issues early. Further refinement of images involves correcting minute details like fingers or object appearances, often using Photoshop for color adjustments to maintain visual consistency.
For animation, Kling was the primary tool due to its effectiveness. The speaker explains tips for complex animations, such as clearly numbering action sequences (e.g., 1. Look, 2. Drink, 3. Smile). It's noted that often only a 'start frame' is needed for animations, rather than both start and end frames, especially when connecting different shots.
The final stage involves editing in Premiere Pro. This includes continuously integrating generated animations, adding sound effects (from Artlist) to enhance realism, and polishing the overall flow and atmosphere. Color grading using plugins like Dehancer and balancing audio levels for background music and narration are also critical steps to achieve a high-quality final product.