Tool Use - AI Conversations Fine-Tune Your Own A.I. Video Model (ft. Greg Schoeninger)
11 snips
Feb 10, 2026 Greg Schoeninger, CEO of Oxen.ai, explains how to fine-tune image and video AI models for cost, consistency, and scale. He covers when to move beyond prompts, dataset creation and labeling, LoRA versus full fine-tuning, segmentation masks, and real-world wins like a massive product catalog and the Isometric NYC project. Practical tips for training, inference, and tooling round out the conversation.
AI Snips
Chapters
Transcript
Episode notes
Generate Training Data Then Fine-Tune
- For massive product catalogs, collect model outputs and curate them as training data instead of relying on API generation at scale.
- Fine-tune an open-source model to cut generation time and cost by an order of magnitude.
Begin With Small, Focused Datasets
- Start with 20–30 high-quality examples to teach a style or character and expand per variation needed.
- Use ~50–100 examples per distinct attribute (e.g., color) so the model generalizes to permutations.
Caption Data To Preserve Controllability
- Caption training items with the exact attributes you want to control later; include color, makeup, or any variable you may change.
- Anything omitted from prompts risks becoming a baked-in default in the fine-tuned model.