Barış Sabuncular · AI Video Producer

Client work

9:16 social ads, delivered

Freelance work through an agency, 2026, made with online tools. Both ads were delivered and approved by the client.

Minicenter: rebrand announcement

Mom & baby retail chain, Baku · Aug 2026 · 36 s · 1080×1920 Instagram

A consistent photoreal AI cast of three, wearing the client's real products, built from their product photos.
Localized on-screen text and music, plus a 31 s alternate cut.
Client-approved after about six revision rounds in eight days.
The edit is built as code (Python + ffmpeg), so timing and text revisions re-rendered in minutes with no new AI generations.

Nano Banana ProKling 3.0ElevenLabsPython + ffmpeg

Fashion ad: the elevator

Clothing brand · Jul 2026 · 12 s · 1080×1920

One consistent AI model across three outfits, from activewear to office wear.
Outfit changes are cut on the elevator doors, so each new look lands as the doors open.
A paid ad for a clothing brand. The brand isn't named at the client's request.

AI fashion modelOutfit consistency9:16 social

Own production · local pipeline

A Diablo story series, made on one 8 GB GPU

A cinematic retelling of the Diablo games for my YouTube channel. Episode 1, "The Butcher" (15 min, about 200 shots), is in production. Every frame, voice and note is generated locally, with no cloud.

Pipeline code ↗

Lazarus, rain. Character LoRA, Wan 2.2, SeedVR2 to 1080p.

The same LoRA in a new set: identity holds across shots.

Moreina, rain. Her own LoRA, same pipeline: test footage before the final grade.

The cast

Each character is defined once, with a written character block plus a trained LoRA where needed, and stays the same person from shot to shot.

72 / 72sound events land within one frame (31 ms) of the picture

17 vs 3of 23 impacts in sync: placed Foley vs AI video-to-audio

31 / 36images on-model with Aidan's LoRA, vs 1 / 36 from the prompt alone (blind AI judges)

8 GBconsumer GPU: keyframes, video, upscale, voice and music, all local

Pipeline: Z-Image keyframes → character LoRAs (musubi-tuner) → Wan 2.2 image-to-video → SeedVR2 upscale + frame interpolation → spotted Foley, Qwen3-TTS voice, ACE-Step music → mix. Every model, music and sound source is checked for commercial licensing.

An unofficial fan retelling. Diablo is a trademark of Blizzard Entertainment. No original game music or voice lines are used.

Open source

Pipelines and tools

I set the architecture, the specs and the acceptance tests, and I build with AI coding agents (Claude Code).

ai-film-pipeline

A local AI film pipeline: keyframes, character LoRAs, image-to-video, upscaling and frame-accurate sound.

game-asset-pipeline

A 3D-to-sprite pipeline with headless Blender rendering and in-engine validation harnesses (Godot).

sprite-qc-agent

An agentic generate, measure, revise loop on the Claude API, gated by a pixel-level QC harness.

vn-art-pipeline

A ComfyUI + SDXL production system: workflow-as-code, automated defect screening, character consistency.

Toolkit

What I work with

Client work runs on professional online tools. My local pipeline is my lab for custom methods such as trained character models and frame-accurate sound.

Online toolsKling 3.0 (video), Nano Banana Pro (image), ElevenLabs (voice, music, sound effects, image and video)

Local, open modelsComfyUI, Wan 2.2, Z-Image, SDXL, Qwen-Image-Edit, SeedVR2, character LoRA training

SoundFrame-accurate Foley and sync, AI voice (ElevenLabs, Qwen3-TTS), AI music (ElevenLabs, ACE-Step)

EngineeringPython, ffmpeg, Claude Code and the Claude API, Blender scripting, Godot