How to Build a Visual Bible for AI Filmmaking: Lessons From SOMOS5's Hybrid Production Workflow
SOMOS5 editor Pablo Díaz explains how his studio uses AI for treatments, previs and hybrid productions. A practical guide to the visual bible, the shot workflow, how to evaluate new models, and where traditional filmmaking still wins.
New AI video models arrive almost every week, and it is tempting to rebuild a workflow around each one. Pablo Díaz, director and editor at the Spanish production company SOMOS5, takes the opposite approach. In a September 28, 2026 interview with LBB, he describes a process where the stable part is not the model but a “visual bible” of locked references, and where AI is used most heavily in pre-production rather than as a replacement for the shoot. This article turns that interview into a practical guide for directors, agencies and small studios.
Where AI helps most: pre-production
Díaz says the biggest impact of AI on his work is in visual treatments, aesthetic development and previsualisation. He can test characters, locations, atmospheres, framing and lighting before hiring a crew. “AI dramatically reduces the distance between an abstract idea and its visual representation,” he told LBB, while stressing that it informs creative decisions rather than replacing them.
That changes the pitch. SOMOS5 used to build treatments from references to other films, photographs and campaigns. Now it generates images specifically for the project, which Díaz describes as turning the treatment into “something closer to a prototype of the film”. The downside is that striking images are now easy for everyone to produce, so differentiation has to come from “the idea, the point of view, the visual coherence” rather than spectacle.
He also uses AI in post-production for clean-up, set extensions and generating specific elements, but sees pre-production as where it changes the most.
What a visual bible is and why it matters
In a traditional production, the art department finds or builds locations, props and costumes. In an AI production, Díaz explains, those elements must first be created and locked as digital assets. The visual bible is the set of approved references that keeps them consistent across shots:
| Section | What to include |
|---|---|
| Characters | Front and three-quarter views, face details, proportions, wardrobe |
| Locations | Wide establishing images, key angles, time of day |
| Materials and props | Textures, surface finishes, hero objects |
| Palette | Primary colours, grading references |
| Lighting | Key and fill direction, quality, practical sources |
| Camera language | Lens feel, framing rules, movement style |
“The more coherent information we give the model, the greater our chances of maintaining continuity,” Díaz says. The bible also makes the workflow more independent of any single model: if a better tool appears, the same references can be fed into it.
For techniques that help keep characters consistent, see our guide to consistent AI character design methods.
The SOMOS5 shot workflow
Díaz describes a sequence that mirrors traditional production:
- Script and shot breakdown. Start from a brief or script and define every shot.
- Look development. Create and approve the characters, locations, wardrobe, art direction and overall aesthetic, forming the visual bible.
- Still frames. Generate a still image for each shot using the approved references.
- Animation. Animate the stills to achieve the camera movement and actions each shot needs.
- Finishing. Edit, sound design and colour grade for the final master.
Generating approved stills before animating is a practical cost control: still images are cheaper and faster to iterate than video, and problems with composition or continuity can be fixed before spending video credits.
Tools: platforms over single models
Díaz’s go-to tools are Magnific and Higgsfield, which he uses “as central platforms for testing different models and combining image generation, video generation, and visual enhancement workflows”. The specific model changes from project to project. Aggregator platforms like these give access to several image and video models through one account, which suits a workflow where the bible stays fixed and the model is chosen per shot. Our article on Magnific One and Brand Kit covers one such platform in detail.
How to evaluate a new model
Rather than chasing every release, Díaz tests new models “through specific exercises” and compares them on:
- control
- consistency
- movement
- speed
- cost
- commercial-use terms
A simple way to apply this is to keep a standard test pack drawn from your visual bible: one character close-up, one wide location shot, one shot with a specific camera move, and one product or prop shot. Run each new model on the same pack and record the results, the number of attempts and the cost. Commercial terms deserve the same attention as quality; a model you cannot legally use for a client’s campaign is not an option, however good it looks.
The recent advances Díaz considers most significant are better fidelity to references, character consistency, camera movement control, image-to-video animation, and the ability to sustain an action for longer. In his words, “we are gradually moving from generation that feels almost like a lottery towards a tool that is beginning to allow for more precise direction.”
Where traditional filmmaking still wins
Díaz is clear about AI’s limits. Its main weaknesses are control, consistency and human performance: models can produce artefacts, alter characters unintentionally and make precise actions hard to specify. He considers AI especially weak at realistic, sustained performance and emotion.
Traditional methods remain faster and more reliable, he says, for:
- long dialogue scenes;
- complex physical actions;
- precise interaction between several characters;
- continuity across many shots;
- completely faithful product representation.
“Sometimes, it is simply easier to place an object on a set and rotate it 45 degrees than to repeatedly explain to a model how it should do it,” he notes. For these reasons SOMOS5 currently favours hybrid productions, using AI to extend environments and generate elements while live action provides performance and precision.
The crew for an AI or hybrid production
Díaz argues the change is not simply hiring someone to write prompts. On a small generative production, the crew might include a director or creative lead, an AI artist for image-making and visual development, a specialist in animation or video generation, a post-production team and a producer, with sound, colour, VFX and legal supervision as needed. In hybrid work, AI should integrate with existing departments: a cinematographer can use it to develop lighting, and a production designer to explore spaces and materials.
Legal and ethical priorities
Asked about legal uncertainty, Díaz highlights concerns about unauthorised use of creative works, generating people’s images or voices without consent, data privacy, and the difficulty of telling real from manipulated content. For a studio, practical steps include documenting consent for any real person’s likeness or voice, checking each platform’s commercial-use terms, and keeping records of which models and references were used for each shot.
For a broader view of the tools, see our comparison of AI video creation tools like Runway and Pika and our report on the debate over AI in anime trailers.
Our analysis
The most transferable idea from SOMOS5’s approach is that the visual bible, not the model, is the core asset of an AI production. It keeps continuity under control, makes it easy to switch models, and turns AI from a “visual slot machine”, as Díaz puts it, into a tool that can be directed. Combined with a hybrid approach that keeps live action for performance and precise action, it offers a realistic way for small studios to use AI without losing the craft that clients pay for.