MiniDog
MiniDog is a lightweight text-to-image diffusion Transformer model trained on consumer-grade hardware (4x RTX 3090). It has only 22 million parameters and supports pretraining from scratch and fine-tuning for generating images of specific dog breeds.
Timeline (last 90 days)
The MiniDog project can complete training of a text-to-image diffusion Transformer on four RTX 3090s: 55 minutes for pre-training and 15 minutes for fine-tuning.
MiniDog's transformer backbone has only 22 million parameters, far smaller than the billion-scale parameter counts typical of mainstream text-to-image models.
MiniDog's pre-training uses 26,000 dog images from ImageNet, and generates text descriptions for them.
MiniDog's fine-tuning uses 2,000 images generated by FLUX.1-schnell, constituting fine-tuning on synthetic data.
MiniDog adopts E2E-INVAE because its semantic latent space carries more interpretable structural information than a conventional VAE.
MiniDog's text encoding uses a small Qwen model, combined with in-context conditioning for conditional injection.
For training acceleration, MiniDog introduces iREPA and the EUPE perceptual encoder.
MiniDog's evaluation covers FID, HPSv2, PickScore, and various VAE reconstruction metrics.
The MiniDog project's code and evaluation tools have been open-sourced.
All Facts (9)
The MiniDog project can complete training of a text-to-image diffusion Transformer on four RTX 3090s: 55 minutes for pre-training and 15 minutes for fine-tuning.
50%UnverifiedMiniDog's transformer backbone has only 22 million parameters, far smaller than the billion-scale parameter counts typical of mainstream text-to-image models.
50%UnverifiedMiniDog's pre-training uses 26,000 dog images from ImageNet, and generates text descriptions for them.
50%UnverifiedMiniDog's fine-tuning uses 2,000 images generated by FLUX.1-schnell, constituting fine-tuning on synthetic data.
50%UnverifiedMiniDog adopts E2E-INVAE because its semantic latent space carries more interpretable structural information than a conventional VAE.
50%UnverifiedMiniDog's text encoding uses a small Qwen model, combined with in-context conditioning for conditional injection.
50%UnverifiedFor training acceleration, MiniDog introduces iREPA and the EUPE perceptual encoder.
50%UnverifiedMiniDog's evaluation covers FID, HPSv2, PickScore, and various VAE reconstruction metrics.
50%UnverifiedThe MiniDog project's code and evaluation tools have been open-sourced.
50%