Training video generation with Wan 2.2: Conan O’Brien and Will Smith character consistency

Links + Notes 📝 https://www.oxen.ai/blog Join Fine-Tune Fridays 🔧 https://oxen.ai/community Discord 🗿 https://discord.com/invite/s3tBEn7Ptg Use Oxen AI 🐂 https://oxen.ai/ Oxen.ai offers one click fine-tuning or fine-tunes for you! Built on top of the worlds best data versioning tool, we offer tools to automate model evals, generate synthetic data, and effortlessly fine-tune models. -- Chapters 0:00 Is it Wan like “Anne” or “won”? 0:55 The Wan suite of models 1:10 Wan 2.1’s model architecture and research paper 3:50 Wan 2.2 video improvements from Wan 2.1 5:35 Our fine-tuning goal: Conan O’Brien interviewing Will Smith who’s wearing a Denver Broncos shirt 7:30 Base model results 8:55 Wan 2.2’s model architecture 12:55 Fine-tuning: How we created our data 17:12 Fine-tuning: How we fine-tuned each Wan model 19:22 Question: How many images do you need? 20:24 Question: Did we use musubi-tuner? 20:40 Question: How to train camera panning 22:45 Fine-tuning: Comparing images as we fine-tune 29:37 Bringing our Will Smith fine-tuned model to Comfyui 42:00 Configuring Comfyui to run our fine-tuned model 47:28 Question: Does the image input format matter? 48:40 Loading our Conan O’Brien fine-tuned model on Comfyui 57:45 Question: How are the LoRAs loaded into the pipeline 58:40 Final Results: Conan interviewing Will Smith