Proprietary IP: Air-Gapped Local AI Production Pipeline
| The Long Walk DIT — 2025 |
Tales from the Loop (8 eps) DIT — 2020 |
NOBODY DIT — 2021 |
| Fractured DIT — 2019 |
The Grudge DIT — 2019 |
Violent Night DIT — 2022 |
|
freeWispur, remote voice dictation
laptop mic · LAN relay · Mac Studio |
Reverse on-set grades back to camera LOG for portfolio comparison sliders.
Alexa · Alexa 35 · Sony Venice |
To ensure absolute data compliance for enterprise studios operating under strict NDAs, I engineer 100% offline, air-gapped automation pipelines that process high-value visual assets locally without internet or cloud-API dependencies.
| Modality | Capabilities |
|---|---|
| Text → Image | T2I generation, iterative refine, prompt-relay chaining |
| Text → Video | T2V with timeline-based prompt evolution |
| Image → Video | I2V with identity preservation, low-noise & high-noise variants |
| Video → Video | V2V style transfer, scene manipulation |
| Video Extension | Temporal extension beyond original frame count |
| Text → Speech | Qwen3-TTS full pipeline — voice cloning, multi-character dialogue, save/load voice profiles |
| Voice Cloning | Clone from short sample → generate speech → feed into Wan Multitalk for talking-head video |
| Lip Syncing | Audio-driven lip sync for character dialogue scenes |
| Character Sheets | Multi-pose consistent character generation via chained ControlNet + PromptRelay workflows |
| LoRA Training | Local LoRA creation and fine-tuning for custom characters and styles |
| Scene Manipulation | Character replacement, background replacement, lighting alteration, compositing — via SAIL2, Bernini, LTX 2.3, Bindweave, Sulphur2 |
Image Backbone:
- 10Eros v1 (FLUX-based, 43 GB) — primary image backbone, custom fine-tuned for all character/genre work
- FLUX.1-dev variants (Kontext, Klein9B) — secondary image backbone
- Z-Image-Turbo (11.5 GB) — real-time turbo generation for rapid iteration
- Qwen-Image-Lightning (1.6 GB) — ultra-fast 4–8 step inference
- SDXL ControlNet suite — OpenPose, depth, Canny for spatial conditioning
Video Generation:
- Wan 2.2 — I2V (low/high noise), Animate (17.1 GB), SCAIL (16.5 GB) for structured video control
- Wan 2.1 — I2V-480P (15.8 GB), Multitalk (2.5 GB) for talking-head integration
- LTX 2.3 — 22B distilled + 1.3B control transformer for efficient high-quality video
- SEEDVR2 (15.4 GB) — video restoration/enhancement
Scene Manipulation & Compositing:
- SAIL2 — structured scene manipulation for character/background replacement
- Bernini — generative scene compositing and lighting alteration
- Bindweave — multi-frame compositing and seam-aware blending
- Sulphur2 — high-fidelity scene reconstruction and manipulation
Audio & Voice:
- Qwen3-TTS (~12 GB total) — full voice cloning pipeline with multiple checkpoints: TTS backbone (3.6 GB) + voice adapter modules (0.65 GB). Supports sample→clone→save→generate workflow and multi-character dialogue.
Key Differentiator Nodes:
- Prompt Relay — chains multiple prompts across a video timeline for smooth narrative transitions. Critical for LTX 2.3 + 10Eros timeline videos and Wan 2.2 structured multi-prompt generation.
- ControlNet Union — combined pose/depth/Canny conditioning for intentional composition and frame-consistent output.
- Qwen3-TTS Voice Pipeline — clone a voice from a short sample, generate speech, save voice profiles, run multi-character dialogues — fully local, no cloud API.
Voice Sample ──► Qwen3-TTS Clone ──► Wan Multitalk ──► Talking Head Video
│
Character Sheet ──► Prompt Relay ──► LTX / Wan Timeline ──┘
│
ControlNets ───────────────────────────────────────────────┘──► Pose / Composition Control
| Workstation | Role | Specs |
|---|---|---|
| Primary AI Node | Air-gapped inference & pipeline automation | NVIDIA RTX 5090 (32 GB GDDR7), Intel i9-14900K, 96 GB DDR5, ComfyUI on local loop (127.0.0.1:8188) |
| Color & Post AI | DaVinci Resolve grading, finishing, AI-assisted post | Apple Mac Studio M2 Ultra (128 GB unified RAM) — air-gappable on demand |
- VRAM Overhead Optimization: The RTX 5090's 32 GB of Blackwell-architecture VRAM enables sub-second local iterations, massive batch arrays, and ultra-high-resolution canvas generation entirely offline.
- Zero-Trust Security Loop: Hardcoded to execute on isolated local host loops (
http://127.0.0.1:8188). Absolutely no telemetry data, client visual payloads, or proprietary training assets cross external network boundaries.
Full Workstation Spec Sheet
| Component | Specification |
|---|---|
| CPU | Intel Core i9-14900K — 24 cores (8P+16E) / 32 threads, boost up to 6.0 GHz |
| RAM | 96 GB DDR5 (2× 48 GB Corsair @ 4800 MHz, rated 6000 MHz) |
| GPU | NVIDIA GeForce RTX 5090 — 32 GB GDDR7 (600 W, 2685 MHz core / 14001 MHz memory) |
| Storage (Internal) | 1.86 TB NVMe (OS + fast cache), 2.8 TB HDD (model storage) |
| Storage (DIT RAIDs) | 9 TB RAID 5 (archive), 32 TB RAID 0 (on-set acquisition) |
| Transport | 5 × 6 TB NVMe RAID shuttle drives (secure footage transport) |
| OS | Windows 11 IoT Enterprise LTSC (64-bit) |
| Software Stack | Python 3.11.4, ComfyUI (local port 8188), Git, huggingface-cli |
| Network | Air-gapped local LAN — zero external network dependency |
| Thermals | GPU idle 34°C / 72 W — max 600 W under full load |
Key Air-Gap Notes:
- No cloud APIs required — all inference runs on the RTX 5090's 32 GB VRAM (with CPU offloading for larger models)
- Entirely local network — zero external service dependency
- 96 GB DDR5 + 32 GB VRAM enables Llama-70B, FLUX, and Wan 2.2 to run comfortably with CPU offloading
- 1.86 TB NVMe (OS + fast cache) + 2.8 TB HDD + 9 TB RAID 5 for model & archive storage
- 32 TB RAID 0 for on-set acquisition, 5x 6 TB NVMe RAID shuttle drives for secure transport
|
After nearly two decades managing complex image pipelines on major film sets for Universal Pictures, Netflix, Amazon, Sony, NBC, CBS, Technicolor, and Deluxe — I now build custom ComfyUI automations and AI solutions for remote teams. I bridge the gap between cinema-grade production discipline and cutting-edge generative AI to deliver pipelines that actually scale in the real world.
São Paulo, Brazil / Manitoba, Canada |
|
ComfyUI Workflows Custom node development, pipeline automation, production-ready image/video generation |
Pipeline Optimization From prompt to output — optimizing for speed, quality, and reliability at scale |
AI Film Services AI colorization, restoration, image-to-video, voice synthesis for production |
B2B Consulting Workflow design, technical advising, remote team integration |
I'm open to teams ready to automate, scale, or debug their AI image pipelines. Let's ship something great.
Crafted with 17 years of cinema pipeline expertise + a dash of AI magic