Spec Ad developed in February 2026
AI CINEMATOGRAPHY
Introduction
AI is Borges' Library of Babel. Every book written and every book that betrayed itself unwritten. Together they are not wisdom. They are its blander, more persistent obstacle. To find art, one does not master the infinite. One refuses it. This word. Not that word. Where babble was... silence.
The noise that was there before and is no longer, and whose absence is the whole point.
Only this. Only now. Only once.
The main objective is that AI does not look like AI. To get there, we have to simulate photography bit by bit. It is not immediate, nor can it be automated. It is the result of rigorous work and disciplined thinking. That said, it does not command the price of a real shoot, nor is it drastically discounted. The investment depends on your needs and the execution timeline.
To ensure both superior output quality and legal defensibility, I have developed a system that works exclusively with hand-crafted elements and plates, operating independently of any generative model or AI system. This guarantees that all production values remain easily curated, revised, and modified. Thereby preserved for future rights claims, regeneration with emerging models, and archival integrity.
This method is described below, though the explanation cannot encompass execution and orchestration, which I reserve for professional discretion. The core advantage is complete ownership of process, heightened control, and accelerated generation timelines. The image emerges from a structured architecture of information clusters. A master "orchestrator" prompt synthesizes all information hierarchically. Detailed guidance follows through keyframe generation, 3D spatial mockups, or actor performance capture.
The workflow begins in Google Vertex, which ensures client data security and supports detailed technical input. Video generation is executed in Luma at the Enterprise tier. This arrangement guarantees your material is excluded from machine learning training, handled as confidential by contract default, and assigns full output ownership to the client with indemnification coverage against third-party claims related to the model's training data.
INTENTION PLATE
I begin with hand-drawn reference sketches that establish the visual intent. I then describe the complete photographic specification to an LLM: exposure, shutter speed, ISO, color temperature, camera body, lens focal length, filters, surface reflectance, film stock or sensor emulation, and specialized photographic technique. The LLM translates this technical language into a precise generation prompt, ordered by hierarchy. I then review, correct, and refine that prompt in detail to ensure every parameter is exact.
Generative models understand photographic language, but only if commanded with precision. Ambiguity in the prompt produces ambiguity in the output. What separates professional results from AI-generated imagery is rigorous specification at every stage.
PENCIL PLATE
Pencil is not better than Photoshop in expressing degrees of light. But it is far better at explaining our
"light intention." There is also the size factor: big drawings let us carefully craft the blocking with exactitude.
One more counterintuitive advantage of this method is that pencil can suggest speed and direction.
DEPTH MAP PLATE
Generative AI with reasoning models attached can be instructed to read depth maps, which are also a form of chiaroscuro created by hand. The advantage of this technique is enormous. Thanks to this step, one can ask the prompt to create shallow focus at a particular distance.
CHARACTER PLATE
Character sheets must be built with meticulous care. The format must be large to preserve fine detail, particularly skin texture and subsurface characteristics. Skin detail is the first element lost in compression and the most critical for photorealism. Without precise reference for micro-texture, light diffusion, and color variation across facial planes, the generated image will read as synthetic, not human.
MACRO TEXTURES PLATE
Models have to know intrinsically what they are doing. Slop is not just a lack of information on our part, but not knowing how much of it to give. The prompt is an orchestrator of elements, not a list to be illustrated.
ENVIRONMENT PLATE
Realistic results depend on how light interacts with the texture of a real place. Everything has to be brought into alignment. Light direction reveals texture. It is much better to align our characters' own textures with the
same source of light and directionality that affects the environment.
SYNTHETIC
PHOTOGRAPHY
Orchestrating all the data I have created is a significant step. The prompt follows hierarchies that sometimes need to be reevaluated. Iteration is unavoidable, so I create a "library of errors," through which I can indicate to the model how close it is to getting it right and what it is missing. I do not use Photoshop for this. If anything, I use red-pen techniques. Then I regenerate the imagery until it reaches 4K quality.
The process can stop here if only still images are needed.
SYNTHETIC FILM
Keyframes, 3D mockups, or filming actors are the main techniques to guide movement. There is no single path to it. However, I do not use animated characters as references. People simply do not move the way 3D animators have learned, imitating Pixar or Dreamworks. To achieve natural movement, I use silences, breaths, or processes of thought within the prompt. If working with an actor, I never "tell them" what to do. I give them context and
help them to believe.
Better than having an actor "acting" is having a human "living."
PROCESS FLOW
THE INVESTMENT
Projects are valued by the days spent on them. I track the time each requires for future budgeting. The process is brand new; the world keeps moving fast and relentlessly. Day rate varies according to a private agreement between you and me.
Factors that affect that rate are: the speed of the exchange, the deadline, the objective of the generation, and whether you are an established company or new to the market. The quicker the exchange, the further ahead the deadline, and the newer the organization, the more economic compression the agreement can accommodate.
I give emerging companies an advantage as I know how hard it is to start.
HARD VS EASY
AI works through latent spaces: fields of possibility encoded in mathematics within a model. A cloud of data tagged with a name.
"Blue" is a latent space of everything that can be blue or mean blue. "Tiger" is another. "Jumping" is another. "Blue tiger jumping" already combines three latent spaces. The more variables we mix, the less stable the combination becomes, especially when we use multipliers. "Seven blue tigers jumping" is one thing. But if we mix it with something unrelated to the subject—say, "Blue Tiger Coffee Jumping"—the model's interpretation demands judgment and discretion. This is the one thing AI cannot maintain stable throughout each iteration.
"Blue tiger made of coffee jumping" would require proper Houdini simulation to be feasible. Everything is feasible with a real team and someone who knows what to call for in each situation. Understanding the limits of the technique helps clarify the budget. Commercials have maintained budget for VFX for years. Demanding that everything be done with AI is risky, to say the least.
Ease comes from alignment and strategic humility. AI generation is already saving us a shoot, but it means nothing if it looks like AI. Easy means solving one challenge at a time. In truth, normality and continuity are also challenges. The goal is to be perceived as real, THEN be fantastic. Layer over layer, one latent space at a time.
HOW DO I DEVELOP THE SHOTS AS
AN AI DOP
My services build on one another. The value lies in conducting a rough session of your vision. Rough sessions are described in the Storyboarding section.
For starters, having a clear idea of the visual narrative is step one. Executing this by hand is what distinguishes me from other artists performing the same work. It is fast, and it provides a safety net against idle iterations where the generated content simply doesn't merit production.
One minute of film would typically require the following timeline.
The first day would be used to conduct rough sessions and determine which plates we need. Then I would spend a couple of days creating large-scale drawings of the entire film. Another day would be for depth maps and spatial referencing, another for character sheets and macro textures. Depending on the project's needs, I would move to Blender to construct specific blocking. This is critical for particular lens choices.
The second week I would dedicate to photography generation and generation refinement. This constitutes the synthetic photography phase.
By the third week we would begin keyframing to develop movement. We can also choose to shoot an actor or produce 3D mockups. It all depends on the content being created.
The fourth week is for video generation and video refinement. I can blend videos seamlessly in DaVinci Resolve if the model requires additional attention. Color correction can also be executed in this week or at the beginning of week five.
THE IP CLEARANCE
for AI CINEMATOGRAPHY
Generative AI does not bypass intellectual property law. It complicates responsibility when something goes wrong, and that liability falls on the agency and the brand, not on the software. My process is built around each of those exposures.
Strict infringement liability
If a model outputs something confusingly similar to existing copyrighted work, the agency and the brand carry the liability, not the AI provider. Ignorance of what a model was trained on is not a defense. Transparency must be planned from the start to build a defensible position for any future legal challenge.
The no-authorship trap
Purely AI-generated output cannot be copyrighted: protection requires a human author with documented creative control, a rule the US Supreme Court left standing in March 2026 and that French and EU courts apply under the same logic. This means all purely AI-generated output cannot be fully owned.
Confidentiality and prompt leaks
Feeding unreleased scripts or new product schematics into public AI generators can
breach an NDA on its own, since those tools may use what you submit to improve their models. I do not work with public-tier tools for this reason. Video generation runs on Luma
at the Enterprise tier, contractually prohibited from training on anything submitted.
Image generation runs on Google Vertex AI, where service terms exclude customer
prompts, files, and outputs from model training. Your material stays yours.
Dossier handoff
What I deliver is not just the final generation. It is the full build nested in files per shot. Every sketch, plate, and 3D file that shaped the image, with the documentation of how each was created. If authorship is ever questioned, you will have a solid defense, traceable from the very beginning of the project.
Curiously, what makes the legal defense strongest is what, in the end, makes AI not look like AI. Intentionality is the defense in court and in art.
HYBRID CINEMA
AI's best space for development is hybrid cinema. To dress the actors and increase the possibilities of the set. Merging seamlessly with effects. Enlarging what was possible until this day. 3D characters have never replaced actors, though somebody tried. AI will go back and forth until its use as a tool becomes established in our global society. The best results come from mixing worlds until a new one becomes the only thing you see.