Field Notes
Exploring what becomes possible.
Small experiments exploring technologies before they become products. Every experiment answers one question: is this genuinely useful?
Local AI · Voice · Video · 3D · Realtime · Agents · Robotics · Game Dev · Creative Tools · Local AI · Voice · Video · 3D · Realtime · Agents · Robotics · Game Dev · Creative Tools · Local AI · Voice · Video · 3D · Realtime · Agents · Robotics · Game Dev · Creative Tools · Local AI · Voice · Video · 3D · Realtime · Agents · Robotics · Game Dev · Creative Tools · Local AI · Voice · Video · 3D · Realtime · Agents · Robotics · Game Dev · Creative Tools · Local AI · Voice · Video · 3D · Realtime · Agents · Robotics · Game Dev · Creative Tools ·
After · locked character sheet
Before · raw generation
↔
Consistent characters across hundreds of generations
Exploring whether modern image models can create production-ready game assets instead of isolated concept art.
Experiment
Testing whether local AI can replace traditional voice ordering with natural conversations and human escalation.Experiment
Swappable parts that stay on-model — outfits, expressions and props recombined without breaking consistency.Experiment
Building a workflow from concept art through Blender into Unity and Unreal using modern generative tools.Experiment
Measuring what belongs on-device, what belongs in the cloud, and where hybrid architectures make sense.Experiment
Building premium launch videos using generative image models, LTX Video and cinematic camera movement.EXP 07 — LTX video experiments
Long-form AI video is finally becoming useful.
▶
Video poster · replace with stillVideo not yet addedHow it was made
Every experiment begins with a question. The interesting part isn’t the model—it’s what survives contact with reality.
01 load reference sheet → face crop
02 ip-adapter · weight 0.62
03 character LoRA · 1400 steps
04 prompt schema: [pose] [outfit] [lighting]
05 seed sweep 8 → pick 1, upscale ×2
06 manual pass: hands, silhouette, palette
Nothing here is exotic. The value is in the constraints: a fixed prompt schema, a small trained model, and a human pass that only fixes the four things the model reliably gets wrong.
Experiment log
Recent experiments01AI-native voice orderingWhisper · Realtime voiceDemo ready02Consistent game character pipelineComfyUI · Flux · LoRA training · PhotoshopWritten up03Cinematic website fly-through generationBlender · SDXLRunning04Local inference architectureOllama · Llama · ClaudeWritten up05LTX Video workflow evaluationLTX · ComfyUIRunning