Image & Video

What's the best open source model for fintuning a large dataset (100k images) of high resolution?

Finding the ideal model to handle fabric textures without losing aesthetic quality...

Deep Dive

A Reddit user with a dataset of 100K high-resolution (2K+) fashion images is exploring open-source models for finetuning, preferring the Apache license to avoid licensing issues. They are evaluating three models: Qwen-Image-2512, ZIB, and ZIT. Key concerns include which model best preserves fabric textures, draping, and aesthetic quality after heavy finetuning, and whether anyone has successfully pushed 100K+ images through these models without catastrophic forgetting.

They also debate full-parameter finetuning versus LoRA for this scale, asking about training cost efficiency versus output quality. The goal is a high-end Vogue look—avoiding the plastic AI aesthetic. The community is invited to share real-world experience on stability, detail retention, and alternative SOTA models that could outperform the listed options. The post highlights the practical challenges of scaling generative models for commercial fashion imagery.

Key Points
  • User evaluating Qwen-Image-2512, ZIB, and ZIT for fabric detail retention after heavy finetuning on 100K high-res images.
  • Debate between full-parameter finetuning and LoRA for 100K samples, with questions about catastrophic forgetting and efficiency.
  • Goal is high-end Vogue aesthetic, rejecting plastic AI look; seeks cost-effective training strategies.

Why It Matters

Real-world insights for generative AI professionals training large-scale fashion models on open-source architectures.