Recruiter Brief.
Detailed answers to the ten questions a hiring team would ask. This page is accessible through a small professional-information link: not part of the main navigation.
Every film starts with a screenplay: scene headings, action, dialogue. From there: treatment, character bible, location bible, storyboard, shot plan. Only then does generation begin. The model is conditioned on references, boundary frames, and identity locks. Post-production is editing, color, and sound. The screenplay is the first and final authority; generation serves the script.
Three techniques in combination: LoRA-trained character models for identity, IP-Adapter / reference conditioning for per-shot anchoring, and boundary-frame reuse: the last frame of Shot A becomes the first-frame anchor of Shot B. The Guitarist case study demonstrates this A→B→C chain in practice. It improves continuity materially but does not guarantee perfection; drift is still possible and is caught in validation.
Continuity is engineered, not wished for. I reuse boundary frames between shots, condition each generation on the previous shot's final state, and validate at the cut point before a segment enters the edit. Where drift is detected, I regenerate with tighter conditioning rather than trying to fix it in post. The principle: the cut is the most visible place for error, so it gets the most engineering.
For images: FLUX and Qwen-Image with LoRA for character-locked generation. For video: Wan 2.2, MiniMax H3, and the LTX family depending on the motion and duration needed. Image-to-video for animating storyboard frames; video-to-video for motion transfer. The choice is driven by the shot's requirements: temporal coherence, motion control, and identity fidelity: not by brand preference.
The gap is post-production. DaVinci Resolve for color and finishing, After Effects for compositing, and Homer Studio (an original tool) for audio. AI output is raw material, not a finished product. Editing, pacing, sound design, and color are where the film is actually made: the same as traditional filmmaking. The model gives me takes; the edit gives me the film.
The Guitarist continuity study. Three independent shots of the same character kept drifting: different face, different clothing, different guitar. I solved it by chaining boundary frames: Shot A's last frame conditioned Shot B's first frame, which conditioned Shot C's. The result is a sequence that holds identity across cuts without perfecting any single generation. The lesson: continuity is a system, not a setting.
AI is a production tool, not a creative authority. It is a camera, a crew, a pipeline. The director's job: intent, selection, pacing, taste: is unchanged. What changes is the cost and speed of producing a take. That democratizes production without democratizing authorship. The films that will matter are still the ones with a screenplay worth shooting.
Custom workflows with Python custom nodes, tailored to each project's needs. I build for reusability: a character-consistency graph becomes a template; a continuity graph becomes a template. I use the Python API for automation and MCP / agent-driven tools for orchestration. The goal is a repeatable process, not a one-off prompt. Where a real ComfyUI JSON exists, it can be mapped into the visual demonstration on the Workflow page.
Local NVIDIA CUDA for training and heavy generation; Apple Silicon / MPS for lighter work and iteration; cloud GPUs for scale. I work with GGUF and FP16/BF16 formats and optimize for memory because constraints are real. The Auteur tool automates production orchestration across these environments.
Work where the screenplay matters as much as the technology. I'm a director who happens to use AI, not an AI technician who happens to make films. I'm most useful where there's a story worth telling and a team that understands that generative production is a craft, not a shortcut.