Introduction
The use of AI in video generation is evolving fast, and we’re starting to imagine workflows where intelligent agents collaborate with human creators. But how realistic is it to automate creative video production using multiple agents? Let’s explore the idea and its potential.
The Vision: Collaborative Creation Between Humans and AI
In a traditional video production workflow, each phase, from the storyboard to the final edit, requires both creative and technical decisions.
AI agents could enhance this process by handling specific tasks, optimizing time, and allowing creators to focus on storytelling and visual direction.
Here’s what an agent-based video creation workflow could look like:
1️⃣ Storyboard and Script Generation
An agent could analyze a concept, theme, or prompt and automatically generate a first draft of the script or storyboard.
Instead of starting from scratch, creators receive structured ideas, visual references, and scene breakdowns to refine manually.
Example:
“Imagine a short film about a robot and a cat exploring an abandoned city.”
→ The AI agent drafts a sequence of key moments, camera angles, and emotional beats.
2️⃣ Image and Animation Generation
Once the storyboard is ready, another agent could handle image or animation generation using diffusion models or video synthesis tools.
This agent could test multiple styles, produce variations, and evaluate visual coherence.
The human role: choosing the most compelling render and adjusting artistic direction.
3️⃣ Post-Processing and Style Consistency
A dedicated agent could analyze all generated scenes to ensure color, lighting, and texture consistency.
It could apply filters or fine-tune outputs to maintain a unified artistic signature across the video.
4️⃣ Automated Editing and Sequencing
The next logical step is an editing agent, capable of assembling generated clips according to the storyboard, synchronizing transitions, and even suggesting background music or ambient effects.
While full automation is still challenging, partial assistance here can save hours of manual editing.
5️⃣ Feedback and Iteration Loop
Finally, a “review agent” could assess the entire production based on the original concept and suggest improvements.
This creates a dynamic feedback loop between human and AI, pushing the final result closer to the creator’s vision.
The Challenges Ahead
- Context understanding: Each agent must interpret the creative intention and coordinate smoothly with others.
- Artistic quality: Automation can accelerate production, but it still struggles to replicate nuanced human creativity.
- Technical complexity: Orchestrating multiple AI systems, tools, and data flows requires robust architecture and synchronization.
A Realistic Approach
For now, the most efficient path is semi-automation:
letting agents assist in brainstorming, visual generation, or post-processing, while the human retains creative control.
This hybrid model leverages the best of both worlds : AI speed and human sensitivity.
Bridging Script and Image Generation: Managing Randomness and Style Consistency
One of the most unpredictable stages in AI video creation lies between the script and the image or animation generation.
Even with carefully written prompts, the results often vary, sometimes subtly, sometimes drastically, due to the internal randomness (seed) of the generation models, which isn’t always reproducible or exposed to the user.
This randomness can lead to surprising creative outcomes… but also to inconsistencies that make it difficult to maintain a coherent visual language across an entire sequence or film.
The Role of Human Sensitivity in Selecting AI-Generated “Rushes”
Each generated frame or animation can be seen as a “rush”, a raw material for storytelling.
Selecting the right ones isn’t just a matter of technical quality.
It requires human sensitivity: understanding the emotional tone, rhythm, and meaning of each shot in the context of the story.
No AI currently captures the intentional emotion behind a scene the way a human creator can.
That’s why this selection process remains a crucial human-in-the-loop phase, where artistic vision guides the narrative cohesion.
Adding a Reference & Pre-Processing Layer
To reduce the chaos of random outputs and keep visual coherence, it’s helpful to introduce an intermediate stage between “Script” and “Image Generation”:
Reference & Style Pre-Processing
In this step, the system uses:
- Reference graphics or screenshots (e.g., from 3D renders or moodboards)
- Multiple perspective samples of the same scene to enforce stylistic consistency
- A pre-processing module that harmonizes visual parameters (lighting, color palette, framing cues) before generation
This ensures that, even if each image is generated independently, they all share the same visual DNA, helping maintain continuity across the video.
Updated Workflow
You can represent this refinement in the workflow as:
Practical Examples: The ΛIVORYA Experiments
To better illustrate how this workflow can be applied in real-world creative projects, here are four experimental videos I created under the ΛIVORYA series.
Each one explores a different aspect of AI-assisted storytelling, combining emotional intention, stylistic exploration, and iterative refinement.
ΛIVORYA | “Tristesse” – Zaho De Sagazan (Fan Art IA)
My first AI-driven video experiment, combining lipsync AI and reference image-based animation.
The main challenge was achieving realistic tear animation and a comic-inspired visual effect that still carried emotional weight.
It involved extensive trial and error to balance realism with expressive stylization.
ΛIVORYA | “Sans Rancune – Becoming Light”
This second project used a single, carefully selected reference image as the starting point.
The goal was to explore different emotional tones and generate expressive rushes through highly detailed prompts.
The focus was on emotional resonance, how variations in AI interpretation could reflect subtle mood shifts in light, expression, and motion.
ΛIVORYA | “Le Cœur du Temps”
Here, I worked mainly with reference images for backgrounds, using them as stylistic anchors for each scene.
One of the biggest challenges was maintaining robot design coherence between sequences, especially across multiple AI generations.
This led me to develop and test a pre-processing approach for visual alignment : an essential step for future projects.
ΛIVORYA | “The Last Link”
This fourth experiment explored an anime/manga visual direction, with highly specific prompts to generate and animate complex sequences, such as the dome destruction and “tree magic” scene in the forest.
After this project, I realized the importance of integrating consistent reference imagery and a style normalization stage before generation to maintain narrative and visual unity.
Reflections on the Process
Through these projects, I learned that AI video creation is more an iterative dialogue than a deterministic process.
Each generated rush (whether successful or not ) helps refine the story’s emotion and coherence.
AI offers endless creative variation, but emotional authenticity still depends on human sensitivity: the way we select, sequence, and emotionally interpret each fragment.
Conclusion
Creating a video entirely through AI agents is still a complex challenge. Yet each experiment teaches us how machines can extend our creative reach.
When emotion guides the process, and technology follows, imagination finds new forms of expression.
That’s the philosophy behind Imalogic : logic serving imagination.
Written by David Lovera, exploring the intersection of AI, creativity, and video production.