I come from film. Executive producer on a feature that premiered at Sitges and played in US cinemas, and VFX artist on a feature currently in post. I bring that craft to generated imagery, because what makes most AI video unusable is not the quality of the single shot: it is continuity, rhythm, and knowing when to stop.
A model generates the shot. Craft is what holds the film together.
Generative VFX artist · Lilicar Films · feature film in post-production
Period thriller directed by Carlos Martín, shot over four weeks on real locations with ARRI Alexa Mini and Zeiss lenses, much of it lit by candlelight. My work is the creature shots: generated elements that have to sit inside footage already filmed and hold up in the cut beside it.
Unit stills published by Camera&Light Magazine · Lilicar Films. The creature shots are under embargo until release.
Cloud video models refuse, on content policy, much of what a genre thriller needs: blood, wounds, animal violence. This is not a technical limitation but a terms-of-service restriction, and it admits no exception and no negotiation. For a film like this one it invalidates the entire workflow: you can generate the animal, but not the scene in which that animal has done something.
The practical consequence is that the work could not be outsourced to a platform. It had to come in-house.
The whole pipeline runs locally, on ComfyUI and an RTX 4090, with open models. That returns three things the cloud cannot give: control of the content; control of the model version — which in the cloud changes without notice and breaks your continuity mid-project; and genuine confidentiality, which on an unreleased feature is not an administrative detail.
On that base, four decisions hold the work up:
This is the commission that proves the method holds inside someone else's production, with their deadlines, their supervision and a cinematography I did not decide.
Three self-authored pieces: brand, script, direction and edit are mine, with no client behind them. They are here because they let me show the whole process without reservation, including what did not work.
The decisions that make a sequence hold — how long a shot runs, when to cut, what stays out of frame — do not come from the model. They come from having been on a shoot.
Feature film
Executive producer and production coordinator on “Everyone Will Burn”. World premiere at Sitges, awarded on two continents, released in US cinemas by Drafthouse.
VFX in progress
VFX artist on Bajo la noche, a feature currently in post for Lilicar Films.
Ten years on set
Between Los Angeles and Madrid: feature film, commercials, branded content and music video. Shooting schedules, budgets and locations.
Post in-house
Editing, compositing and grading in DaVinci Resolve, with the timeline compiled by script from a declarative edit list.
Self-initiated · short film based on the Benito Pérez Galdós novel
62 shots and 7 sequences with a visual reference bible and a continuity protocol: same ships, same characters, same light, shot after shot. Trailer locked; teaser in the mix. This is the piece where the protocol was tested at film scale rather than spot scale.
Finished piece · eight shots
The same couple and the same story carried across different animation styles, from studio 3D to illustration, without losing character identity. The opposite exercise to Trafalgar: there the continuity is of a world; here it is of a character across worlds that change.
Self-initiated · in development
A comic-epic thriller with the whole of pre-production done: dossier, character bible and generation breakdown all locked before the first shot was generated. Teaser in production.
The full protocol told through a real case, with the storyboard, the character sheet, the model comparisons and the post-mortem of the day a whole batch was lost. Not only the flattering part.
See the method on a real case · in Spanish| Character continuity | LoRAs trained on hand-curated datasets, five-view character sheets, ControlNet and IP-Adapter as anchors. The character is an asset of the project, not an accident of the prompt. |
|---|---|
| Image-to-video workflows | First frame, first and last frame, keyframe i2v and multi-ControlNet v2v. Custom ComfyUI workflows with custom nodes, running locally on an RTX 4090. |
| Correction and recovery | Inpainting, figure-and-ground separation, character reorientation from the sheet, and lip sync. A failed shot is regenerated against the correct anchor; it is not patched. |
| Models | Wan, LTX, Flux, Qwen Image Edit, MiniMax and Seedance locally; Kling, Runway, Veo and GPT Image in the cloud. The choice is settled by comparing the same shot, not by reading specifications. |
| Finishing | Interpolation, upscaling, compositing, rotoscoping and grading in DaVinci Resolve. Sound design and mix with loudness metering. The generated shot ends up being an edited shot. |
| Pipelines for teams | I build this system for professionals already using AI who need repeatable results: ComfyUI workflows, bespoke LoRAs, edit automation and training. |
Knowing where not to put AI is part of the job. I would rather say it before signing than after delivering.
Seven finished pieces, to watch back to back.
Enter the screening roomAn AI piece that doesn't fall apart by the second shot?
Get in touch