Practice
Generative-AI tooling that integrates cleanly with existing VFX and animation pipelines — Nuke, Houdini, Maya, Blender — without disrupting the production.
How it works
We design for VFX supervisors and pipeline technical directors. The brief is consistent: generative-AI tooling that integrates with the DCCs, render farms, and colour pipelines already in production. ComfyUI workflows orchestrated on Deadline / OpenCue render farms. Seeded, deterministic generation with locked model versions per show. USD-native asset exchange. ACES colour management throughout.
Inpainting, outpainting, rotoscoping assistance, plate prep, beauty pass cleanup, sky and crowd replacement. Run as Nuke gizmos, Houdini HDAs, or batch jobs on the render farm.
Custom diffusion pipelines with motion-vector conditioning, depth and optical-flow priors, and reference-image fidelity. For element gen, look-dev shortcuts, and previz that doesn't fight the supervisor's notes.
ComfyUI workflow registries, render-farm integration (Deadline, OpenCue, Tractor), USD-aware asset libraries, versioned-asset review tooling, and per-show config management.
LoRA training on a show's reference library, style-locked adapters for series consistency, and per-asset fine-tuning so generative outputs stay on-model across thousands of frames.
Generative tooling that does not integrate cleanly with the show's pipeline — versioning, colour management, review, archival — is not delivered. We engineer to the requirements of the pipeline, not to the demonstration.
Diffusion models live in sRGB; production lives in ACES. We engineer the colour transforms end-to-end so generative outputs sit correctly in scene-referred space, not the supervisor's inbox.
Seeded generation, locked prompts, frozen model versions per show, and per-shot config files in the production repo. Reproducibility is a delivery requirement, not a feature request.
Generative jobs queue on Deadline / OpenCue / Tractor like any other render task. No bespoke infra to babysit. No separate cost-tracking system.
Model training data and asset provenance tracked from day one. No "trust us" on what went into the model.
If we can't verify what went into a model, the studio can't ship outputs from it without legal exposure. Provenance is engineered in from the start.
Recreating a specific performer's likeness requires their written consent and a per-use contract. We won't build the pipeline that lets a studio skip that.
Every generative tool we ship comes with an eval suite for the show — pass rate, supervisor accept rate, rework rate. If we can't measure it, the show shouldn't bet on it.
Tell us what you've already tried, what you've ruled out, and what success looks like. We come back within one working day.
