Generative Artist — Image & Video (Contract)

Designing Digital Products is Never Easy. Power Shifter Can Help.

About The Position

You generate finished shots — stills and motion — to a locked spec. Directors set intent and deliver first-pass keyframes. You take those keys and produce the volume: variants, coverage, alternates, the shots that fill the cut. Editors assemble. You do not edit. This is a craft role with a production spine. The studio runs on locked references and dated sign-offs, and you generate against what is locked, never against a script or a vibe.

Requirements

Interpreting direction. You can take a keyframe, a reference, and three sentences from a director and return something that reads as the same film. Taking a note without defending the first attempt is the job.
Model fluency across the field. Deep working knowledge of current image and video models and a demonstrated history of picking up new ones fast. We do not care which tools you list today — the stack turns over every quarter. We care that you know which model to reach for and why, and that you were fluent in a model that did not exist twelve months ago.
Consistency under pressure. Character and set persistence across shots is the hardest problem in this pipeline and the one that destroys budgets. Reference sheets, seeds, character references, image-to-image chains, control conditioning — you should have opinions about all of it.
LLM prompting as a working tool. Prompt construction, prompt libraries, structured prompt data, using an LLM to expand and vary shot prompts at scale. Claude as production infrastructure. This is not a bonus skill here — it is how the studio runs. You should be working in Claude daily and be able to Use and extend our existing Claude Skills — the shot-list, prompting and estimating skills the pipeline already depends on
Author a new Skill when a repeatable task shows up twice: write the instructions, structure the reference files, test it, hand it to the team
Connect and operate MCP servers to wire Claude into the tools you use — generation platforms, asset storage, project tracking — and know the difference between a tool a model can call and a prompt that describes one
Judge when an agentic workflow genuinely beats doing it by hand, and say so when it doesn't
You do not need to be an engineer. You do need to be someone who automates their own job rather than repeating it.
AI voice and sync. Voice generation, cloning within licensed bounds, lip-sync tooling, and the judgment to know when a line needs to be re-written rather than re-synced.
Post-generation craft. Upscaling, interpolation, grain and noise matching, colour continuity across takes, plate cleanup. Enough compositing sense to know when a shot should be fixed in comp instead of regenerated.
Cost and throughput literacy. You know roughly what a shot costs in credits and time before you queue it, and you make the regen-vs-comp-vs-escalate call yourself.
Rigorous file hygiene and versioning. Every asset traceable to its recipe.
Working knowledge of commercial rights and model terms — what can ship in a paid client campaign, what carries training-data risk, what needs disclosure.
Studio accounts only. No client material in personal or consumer-tier accounts, no client data used for training.
Async discipline. Consolidated updates, clear contact sheets, no silent absorption of risk.
Overlap with Pacific Time for review windows.

Nice To Haves

Familiarity with C2PA and EU AI Act transparency obligations is an advantage.
Local pipeline experience (ComfyUI, node graphs, local GPU) alongside hosted tools
Scripting for batch and API work
Having built your own MCP server, agent workflow, or automation for a creative pipeline
Traditional production, VFX, animation or photography background
3D or previs literacy for blocking and camera intent

Responsibilities

Generate image and video shots to QA'd keyframes, style tokens, and reference sheets — matching an existing look rather than inventing one
Hold continuity: characters, sets, props, wardrobe, lighting and lens language consistent across a shot, a sequence, and an episode
Run batch generation and triage contact sheets down to selects
Produce AI voice and lip-sync passes, including alternate line handling when a read fights the model
Log every take with prompt, seed, model, version and parameters — provenance is mandatory, not optional
Flag feasibility before spend: what fights the model, what costs 3x, what should be composited or shot instead
Deliver editorial-ready material to spec — correct codec, frame rate, resolution, colour space and file naming
Escalate a shot that will not land inside its time box instead of grinding on it