HDM Demo

Enter a natural language prompt and/or specific tags to generate an image.

HDM: HomeDiffusion Model Project

HDM is a project to implement a series of generative model that can be pretrained at home.

Usage

This early model used a model trained on anime image set only, so you should expect to see anime style images only in this demo.

For prompting, enter danbooru tag prompt to the box "Tag Prompt" with comma seperated and remove the underscore. enter natural language prompt to the box "Natural Language Prompt" and enter negative prompt to the box "Negative Prompt".

If you don't want to spent so much effort on prompting, try to keep "Enable TIPO" selected.

If you don't want to apply any pre-defined format, unselect "Enable TIPO" and "Enable Format".

Model Spec

  • Backbone: 343M XUT(UViT modified) arch
  • Text Encoder: Qwen3 0.6B (596M)
  • VAE: EQ-SDXL-VAE, an EQ-VAE finetuned sdxl vae.

Pretraining Dataset

  • Danbooru 2023 (latest id around 8M)
  • Pixiv famous artist set
  • some pvc figure photos
0.5 2
-0.5 0.5
-0.5 0.5
1 4
1 50
1 7
0 1
0 1
768 1280