Generative Video Supervisor

Location: Barcelona

Por favor, asegúrese de leer atentamente los siguientes detalles antes de enviar cualquier solicitud.
About us

We are Dragons, a global creative agency working with bold brands to build cultural relevance, connection, and long-term impact through strategy-led creativity.

We bridge digital and creative thinking, developing innovative work across skincare, healthcare, fashion, lifestyle, FMCG, and more. Our international teams in the US and Barcelona deliver results through smart strategy, sharp creativity, and meticulous execution.

About the role

As Generative Video Supervisor, you own how AI-generated and AI-assisted video is made, quality-controlled, and delivered in this agency. You're both the person who can produce premium, realistic results with your own hands and the person who teaches the rest of the team to do it, and to recognize when a shot isn't good enough to air. You'll be judged on shots that survive client review and broadcast QC, not on a showreel of eight-second clips that fall apart at second nine.

We are looking for someone with taste, honest about limits, including your own and the tools', comfortable in ambiguity, since there's no playbook and you'll help write it and able to say "no, we shouldn't" to a client, a creative, or your own boss, with reasons.

Key responsibilities
Produce premium, realistic results

• Create generative video for real deliverables: concept and pre-vis, animatics, B-roll, plates, transitions, social-first content, and extension or cleanup of existing footage
• Work across text-to-video, image-to-video, video-to-video, restyling, extension, and re-framing, choosing the right method per shot instead of defaulting to one tool
• Define what "premium" and "realistic" mean here as concrete acceptance criteria: motion quality, physics, lighting, texture, skin, hands, text, temporal stability, and match to camera footage
• Push output to delivery quality: upscaling, frame interpolation, denoising, grain and texture matching, color matching, and finishing in the edit and grade
• Know where generation fails today and route those shots to conventional VFX, shooting, or stock, without ego

Consistency

• Maintain character, product, wardrobe, environment, and style consistency across shots and sequences
• Use reference-based conditioning, image-to-video from controlled keyframes, structural conditioning (depth, pose, edges), and fine-tuned models or LoRAs where appropriate, trained only on data we have the rights to use
• Design shot-by-shot workflows (keyframe first, then motion, then finishing) instead of gambling on one-shot prompts
• Build consistency checks into review: side-by-side comparison, continuity sign-off, and artifact logs

Reproducibility and versioning

• Establish a reproducibility standard: every approved shot must be traceable to its exact inputs, so we can regenerate, adjust, or extend it later
• Record and version prompts, negative prompts, seeds, model and checkpoint versions, samplers and parameters, conditioning inputs, workflow graphs, software and driver environment, and source assets
• Understand the limits of seeding: what a fixed seed does and doesn't guarantee across model versions, hardware, software updates, and closed APIs that expose limited controls or none
• Build reproducible node-based or scripted workflows (ComfyUI or equivalent) that others can run, with pinned versions and documented dependencies
• Decide, per project, which tools are acceptable given reproducibility, model-version stability, and vendor lock-in risk

Pipelines and integration

• Integrate generation into the editorial pipeline: conform to project frame rate, resolution, color space, and bit depth; deliver as ProRes, image sequences, or EXR as needed
• Connect generation with Premiere Pro, After Effects, and DaVinci Resolve using scripting, ffmpeg, interchange formats (XML, EDL, AAF, OTIO), watch folders, and queues
• Manage compute: local GPU vs. cloud, queueing, cost per usable second, and what may run where for confidentiality reasons
• Build QC for AI output: artifact and flicker checks, spec validation, and mandatory human sign-off before client delivery
• Track cost, time, and hit rate per workflow (usable shots per generation attempt), so we know what's worth keeping

Evaluation and tool watch

• Test video models and tools (commercial and open source) against fixed benchmark briefs, with documented results on quality, consistency, controllability, speed, cost, license terms, and legal risk
• Maintain an internal tool map: what we use, what we tested and dropped, and why
• Track a field that changes monthly, and filter signal from noise for the team

Teach and set standards