MiniDog: Training a Text-to-Image Flow Matching Model on Dog Breeds in Under 70 Minutes

Forum topic · 小凯 · 2026-10-15

Summary

MiniDog is a framework showing that text-to-image diffusion transformer training is accessible without frontier-scale compute. The authors pretrain a 22M-parameter diffusion transformer in 55 minutes and fine-tune it in 15 minutes on four RTX 3090 GPUs. Pretraining uses 26k captioned dog images from ImageNet, while fine-tuning uses 2k images generated by FLUX.1-schnell. Efficiency comes from precomputing and caching both image and text features. The architecture combines E2E-INVAE for its semantic latent space, a compact 22M-parameter transformer, a small Qwen text encoder, and in-context conditioning. Training is accelerated with iREPA and an EUPE-aware encoder. The released repository supports evaluation via FID, HPSv2 score, PickScore, and multiple VAE reconstruction metrics. The work (arXiv:2610.18870) by Esmailizadeh, Leng, Liang, and Zheng targets beginners in computer vision who want hands-on experience training generative models.

Paper Overview

Field: Computer Vision Authors: Reyhaneh Esmailizadeh, Xingjian Leng, Zhanhao Liang, Liang Zheng arXiv: 2610.18870

Abstract

We generally perceive that training text-to-image diffusion models requires heavy compute: it is true if we aim at a comprehensive frontier model. This understanding prevents people from training their own models when getting started in the field. We develop MiniDog, a framework that enables pretraining diffusion transformers in 55 minutes and fine-tuning in another 15 minutes on four 3090 GPUs. For data, we use 26k dog images from ImageNet and caption them for pretraining and use 2k images generated by FLUX.1-schnell for fine-tuning. We precompute the image and text features and cache them for efficiency. For architecture, we use E2E-INVAE for its semantic latent space, a small transformer with 22M parameters, a small Qwen model for text encoding, and in-context conditioning. For training, we use iREPA and an EUPE-aware encoder to accelerate convergence. For evaluation, the repository supports FID, HPSv2 score, PickScore, and multiple VAE reconstruction metrics.

Key Takeaways

Accessible compute budget: full pretraining (55 min) + fine-tuning (15 min) on 4x RTX 3090 GPUs.
Data: 26k captioned dog images from ImageNet for pretraining; 2k FLUX.1-schnell-generated images for fine-tuning.
Efficiency: image and text features are precomputed and cached.
Architecture: E2E-INVAE latent space, 22M-parameter transformer, small Qwen text encoder, in-context conditioning.
Training tricks: iREPA and EUPE-aware encoder for faster convergence.
Evaluation: FID, HPSv2, PickScore, and VAE reconstruction metrics supported in the repo.