AI Video Agency Fully Reveals High-Quality Video Production Workflow

If initial stages such as idea concretization, world-building, and character setting are skipped, and image production begins immediately, maintaining video consistency becomes very difficult. This significantly impacts the final video's completeness.

Only by proceeding step-by-step can high-quality videos that align with the planning intention be stably produced.

A case study of a beauty film internally produced by the AI video agency was introduced. This film, featuring dreamlike storytelling, was created as an advertisement for a body lotion brand. In the planning stage, the concept was set as finding the source of the body lotion's scent by searching for fruit in a maze.

Through this storytelling, the aim was to visually and emotionally convey the product's characteristics. A clear planning intention enhances efficiency in the AI-driven production process.

Specifically, to create a dreamlike atmosphere, reference images were collected focusing on keywords related to sunset glow such as 'sunshine,' 'hazy lighting,' and 'golden light hour.'

Pre-setting these visual keywords greatly helps in giving specific instructions to AI and maintaining a consistent video tone.

World-building and character setting are critical stages that determine the depth and consistency of a video. By providing mood boards to the AI, it extracts professional keywords such as 'nostalgic,' 'soft,' 'sundrenched,' and 'unposed' to assist in the setup process.

Finding mood board references is essential for obtaining these professional keywords and prompting the AI to generate more accurate images. This allows for clear establishment of a world and characters that align with the planning intention.

These bots are utilized from the planning stage to generate detailed virtual product and model images.

When a proposal is input, they create product visuals and staging shots, and even provide 360-degree turnaround images of models, contributing to maintaining consistency throughout the video. This allows for quick review of various drafts without actual filming.

A feature for dressing AI models in various outfits was introduced. By using the 'fashion image generation' function, different styles of clothing, such as business casual and sporty street casual, can be fitted to AI models. This process automates prompt generation, significantly reducing working time.

Previously, the tedious process of requesting prompts from ChatGPT and pasting them into Exfield was required, but now various outfit images can be generated directly, increasing efficiency.

The importance of planning was re-emphasized as the most crucial stage. Neglecting or skipping the planning stage can lead to prolonged unnecessary trial-and-error (colloquially, 'digging holes') during image or video production and excessive consumption of AI tokens.

The AI agency thoroughly conducts these planning tasks using Claude, explaining that solid planning is a key factor in increasing the overall production process efficiency and reducing costs.

It was emphasized that just like in actual video filming, model outfits appropriate for the video concept are very important in AI video production. If the AI is informed of the product concept and asked for suitable model outfits, it generates specific prompts such as 'flowy, delicate summer outfit.'

Such specific prompts play a decisive role in helping the AI create images that align with the planning intention, ultimately contributing to improving the overall quality of the video.

This is because if AI generates a 10-second video, it may start well but its quality might degrade or become boring towards the end.

Therefore, videos are created by generating several short scenes of about 3-5 seconds, then cutting them down to 2-3 seconds. This method maintains the quality of each scene and ensures editing flexibility.

It was explained that obtaining desired results from AI image generation often requires multiple attempts. Since AI frequently adds unexpected elements or generates images different from the planning intention, a continuous process of modification and dialogue is essential.

For example, if a mirror suddenly appears in the middle of a maze, one must correct the AI's error by saying 'Correction.' Through countless trials and persistent communication, the desired image can ultimately be extracted.

It was explained that while videos can be created with a simple 'click' method, expecting high quality from such videos is difficult. To produce high-quality videos like the one shown earlier, each scene must be meticulously worked on, almost like 'artisanal craftsmanship.'

This refers to the process of precisely extracting desired images through extensive dialogue with AI and repetitive prompt generation. It was emphasized that easy and fast methods cannot yield results of commercial quality.

It was stated that Magnific is primarily used in the final video production stage. The main reason for choosing Magnific is its 'Space' feature, which is highly useful for collaboration. Within a Space, one can not only generate images or videos but also write storyboards and record common prompts using memo functions.

Similar to Figma as a collaboration tool, it allows team members to share work, exchange opinions, and efficiently proceed with projects in one shared environment.

It was emphasized that while videos for hobbies can be easily made with a 'click,' creating 'money-making' quality videos that can be paid for requires significant effort and wrestling with AI. The common 'click' method on YouTube is not enough to obtain high-quality results usable in professional settings.

For brand operators creating advertising videos or videos for commercial purposes, videos made with simple methods lack competitiveness. Therefore, delicate work and continuous interaction with AI are essential for professional-level videos.

Through extensive dialogue and prompt generation, desired images must be precisely extracted, and based on this, high-quality videos must be created to generate tangible value.

First, to prepare for server errors that may occur right before deadlines. Relying on only one server carries significant risk, so using at least two AI tools in parallel is more stable.

Second, Exfield is active in service development, releasing new features weekly, and generates natural images well through its proprietary image model 'Soul2'. While there is some switching cost (time and energy) involved in individually downloading images extracted from the model studio and then converting them into video, Exfield has unique advantages.

Third, Magnific's 'Space' feature is highly useful for collaboration. Most tasks, including image generation, video production, and memo writing, can be handled within Space, enhancing communication and work efficiency among team members. If Exfield excels in individual image generation, Magnific excels in providing an integrated collaborative environment.

The necessity of countless trials and errors during the image generation process using Magnific was emphasized. Citing an example of five attempts to get just one best shot, it was explained that AI continuously produces unexpected results until the final selected image is extracted.

This means that the process of selecting the image that best fits the planning intention from various images proposed by AI, and continuously modifying insufficient parts, must be repeated. This highlights the effort and persistence required to obtain high-quality images.

The criteria for choosing between Magnific and Exfield were explained. It was stated that there isn't a huge performance difference between the two tools, but rather convenience, efficiency, and ease of work collaboration are important selection criteria.

It was explained that considering the unique features and workflows offered by each tool, it is important to choose the tool that is more suitable for the nature of the project or the team's working method. This suggests that factors other than technical performance should also be considered when selecting AI tools.

The first step involves discussing with Claude or GPT in the in-house studio linked with Exfield to create virtual products and extract various staging images. Next, Exfield's 'Soul2' model is used to generate natural model images, and consistency is maintained through the turnaround function, with fashion fitting to confirm suitable outfits for the model.

Once the product and model are confirmed, the storyline is developed with Claude. At this point, mood board references obtained from Pinterest etc., are provided to concretize the concept, and the scene-by-scene storyline, from the opening to key scenes, is meticulously constructed.

It was explained that after the product and model are confirmed, storyline development begins with Claude. Providing mood board references at this stage is crucial, presenting dreamlike images found on Pinterest to the AI to concretize the concept.

Among various concepts proposed by AI (options A, B, C), the desired option is selected and used as a basis for developing the storyline. For example, specific scenes are arranged in order, such as the first opening being a mailbox, the second a woman looking at the mailbox, and the third a woman who has read the letter.

After connecting nodes for models, backgrounds, and products within Magnific Space, prompts are input, and images or videos are generated within that space.

This demonstrates a workflow that organically links AI tools to maximize the strengths of each tool. It shows how specific prompts are created with Claude, and then desired images are efficiently generated through Magnific's visualization and collaboration features.

Among the tools used for AI video production, 'Kling' is currently recommended. While 'Siddende' is also a good tool, Kling and Siddende are currently leading the field in video production.

This implies that by using the most effective and stable-performing tools in the rapidly evolving AI video technology landscape, optimal results can be achieved.