I’ve been working with generative AI since 2022. This page traces how that process evolved — from reading academic papers at a startup to delivering campaigns for La Caixa and Orange. It runs from the most recent work back to the beginning.
I built my own production system, and it is almost entirely agentic. At its core is a knowledge base in Obsidian that learns from every generation and every interaction. I crystallise what it learns into scripts and small apps, for my own speed and for the rest of the team.
Claude runs in my terminal with Adobe skills and bridges I wrote for Premiere and After Effects. It orchestrates generation through ComfyUI, with my own custom nodes, plus Magnific, Kling and Seedance. When a job needs little power, ComfyUI runs locally. When it needs more, I send it to RunPod through a simple button panel I made, which lets me pick the GPU without leaving my own machine. Whatever is generated in the cloud lands automatically in my project folder. Nothing reaches me until the agent has checked and corrected its own output.
Every project is traceable. Prompts go automatically into a spec sheet, outputs are filed by model and phase, and Content Credentials are applied, so anyone can see how each piece was made and repeat it.
2026 — client work in production
This year most of my AI work has happened inside real pipelines: client deadlines, brand standards, teams that aren’t built around AI workflows.
The interesting problems have been technical — getting generative output to hold up when the bar is set by a bank or a telco, not by what looks good in a demo.
- La Caixa — AI video artefacts on brand elements; a hybrid compositing solution
- Orange — first 100% AI campaign; pipeline optimisation and hybrid VFX for logos on moving devices
- Species reconstruction — rebuilding an extinct bird through multi-model orchestration
Late 2025 — Patiele music video (in progress)
This is still in progress. The goal is to find out whether a reliable pipeline can be built for a full music video using generative AI — not getting good frames by luck, but something predictable enough to plan a budget and a schedule around.
The process starts with traditional pre-production: script, shot breakdown and art direction locked before touching any generation tool. From there, I trained custom LoRAs for the lead singer, Ama, so her identity would stay consistent across hundreds of different frames. Sets and characters were designed as animation assets, not improvised through prompts. Before generating any video, I built a complete animatic with stills synced to the music — which let me catch rhythm and timing problems at the cheapest possible point. The final animation mixes different models depending on the shot, and everything goes through traditional compositing at the end.
The ten-step methodology I’m developing here comes from a specific frustration: AI generation is only useful if you can predict what you’re going to get. Randomness is a production problem, not a creative advantage.
August 2025 — Abuelas
This was a personal R&D project. The question I was working on: can generative AI be used inside a filmmaking process without taking over the direction?
The answer is yes, but it requires keeping creative decisions entirely in your own hands. I used AI to generate visual material — sets, characters, atmosphere — and made every decision about rhythm, pacing and emotional arc by hand. The tools were a resource, not a director. That distinction seems obvious until you’re editing generated material, because it comes with its own implicit suggestions about what should happen next. Resisting that is a specific skill, and it takes practice.
July 2025 — Mahou, spatial expansion (Serena)
This was the first AI project I did at Serena. The challenge was to rebuild a floating structure out at sea, with boats, and to add new parts to the building that had been filmed. The brief asked to turn a single shot of a small asphalt platform into a massive maritime structure surrounded by water and boats. No location shoot, fixed budget.
I started with a 3D shot in Cinema 4D to lock perspective and scale before touching any generation — if that foundation fails, nothing that comes after will look physically real. From there I used Adobe Firefly to extend the environment, specifically because it respects geometric references well. Integrating the generated environment with the original footage went through Photoshop and Topaz. The animation was done with video generation tools, keeping the brand asset intact at all times.
2022–2023 — the visual consistency lab (Snippet-Myme)
The images below come from the Midjourney research phase — not isolated prompts, but a sustained investigation into whether specific art direction variables could be held constant across different outputs. Same palette logic, same lighting approach, same character identity across different generations.
I worked through different visual registers on purpose. Some of these were stress tests — pushing prompt construction to the point where the model stopped being useful, specifically to map where that limit was.
This is the part that tends to get skipped: in production, knowing where something breaks matters more than knowing what it does well.
What that research produced was an operational understanding of what is actually controllable and what isn’t — which matters more than it seems when you have a real deadline bearing down on you.
2022 — first commercial project (Snippet-Myme)
The first project where AI-generated assets made it into a real delivery started from a concept by Diego Flores. My role was the technical execution: generation, compositing, editing — turning the pieces into something that worked as a finished film. Pedro Bataller did the sound design from a rough demo I cut. Daniel Fondón wrote the copy.
What I remember most about that project is how much manual work was still needed after generation. The model took you only so far. The rest was traditional compositing and frame-by-frame correction. That ratio — generated versus manual — didn’t change as drastically over the following years as many people expected.
Before that (Snippet-Myme)
I was at Snippet — a video startup — when I started getting seriously into generative AI. The company needed to understand where the technology was really heading, not the headline version, so I went to the academic literature. Diffusion models. Latent spaces. How image consistency works mathematically. I compiled what I found into structured reports in Spanish, and eventually those documents were sold to clients and agencies as intelligence briefings. It wasn’t the original plan, but it made sense.
In parallel I was running tests in Midjourney — not to generate pretty images, but to find out whether visual consistency was achievable at all. Whether you could keep a coherent visual universe across multiple outputs, or whether the tool would keep producing variations that made any professional delivery impossible.
Those questions stayed with me and shaped everything that came after.
