Fine-Tuned Qwen-Image-Edit vs Nano-Banana - Generating 1.2 Million Images

Learn how to generate 1.2 million images using fine-tuned Qwen-Image-Edit versus Nano-Banana while dramatically reducing inference costs in this 44-minute technical webinar. Explore the specific task of generating tables and workbenches in different colors, starting with testing Nano-Banana to determine if fine-tuning is necessary. Examine the pricing considerations and evaluation methods for generated image quality, including how to specify desired colors in the generation process. Discover the fine-tuning workflow from dataset preparation through GPU provisioning with Baseten, and observe the training process in real-time. Master inference optimizations and learn how Lightning LoRA can accelerate inference by reducing the number of required steps. Gain practical insights into large-scale image generation, cost optimization strategies, and the decision-making process between using pre-trained models versus fine-tuning for specific use cases.

Syllabus

0:00 Using Qwen-Image-Edit to generate 1.2 million images and cutting inference costs
5:45 The Task: Generating tables and workbenches in different colors
7:30 Testing Nano-Banana first to see if we even need to fine-tune
13:30 The Pricing Dilemma
16:26 Question: How did we evaluate the generated table quality
17:15 Question: How did we pass in the colors we wanted
18:48 How we kicked off the fine-tuning from the dataset
21:31 How Baseten provisions the GPUs to kick off a training job
24:44 What you see while fine-tuning
26:22 The inference optimizations
37:10 Using a Lighting LoRA speed up inference by reducing inference steps
39:26 General Questions