Image Model Training Recipe Released in 24 Hours

Photoroom has released the results of a Text-to-Image model training speedrun conducted over 24 hours using 32 H200 GPUs and a budget of about $1,500. This experiment focused on finding the optimal training combination to maximize performance within limited compute resources.

The key technical features are as follows:

• Direct Training in Pixel Space: Uses the x-prediction formulation to train directly in pixel space without a VAE. This simplifies the architecture and improves efficiency.
• Efficient Token Management: Uses a patch size of 32 and a 256-dimensional bottleneck layer to control sequence length even at high resolutions.
• Optimized Training Schedule: Instead of the conventional progressive resolution-increase approach, it uses a strategy of starting at 512px and fine-tuning to 1024px.
• Use of Perceptual Losses: Since pixels are predicted directly, various perceptual loss functions from classical computer vision can be easily applied.

Photoroom has open-sourced the training code and experimental framework used in this experiment via GitHub (PRX).