HiT
hierarchical diffusion transformer for image generation
2 min readupdated
HiT (Hierarchical Diffusion Transformer) releases official PyTorch checkpoints for a pixel-space diffusion model (a model that synthesizes data through denoising) that groups related image patches into a visual hierarchy and processes most blocks over compact groups. The model handles unconditional generation on ImageNet as well as text-to-image synthesis. Its architecture streamlines computation through this hierarchical representation while maintaining high visual quality. The release includes HiT-XL and HiT-B model variants across different resolutions.
Key features
- Pixel-space Diffusion Transformer with a hierarchical patch structure
- Official PyTorch weights for ImageNet 256×256 and text-to-image 512×512
- Streamlined block processing using compact groups instead of the full patch set
- Strong image synthesis benchmark scores
What you can do
- Use the HiT-XL/16 checkpoint to generate high-quality 256×256 images on ImageNet
- Run the HiT-B/16 model for 512×512 text-to-image generation from prompts
- Integrate this hierarchical patch approach into your own diffusion model research
Example
To run generation, download the corresponding weights file (such as ImageNet256/HiT-XL16_ep600/model.safetensors or T2I512/HiT-B16_sft40k/model.safetensors) and load it into your PyTorch pipeline following the instructions in the official repository.
Read more
Details• ImageNet 256×256 checkpoint (HiT-XL/16, 600 epochs): FID-50K 1.48
• T2I 512×512 checkpoint (HiT-B/16, 40k fine-tuning steps): GenEval 87.78, DPG 85.33, PRISM 58.62, CoRe 45.22, Spatial 52.68
• Text-to-image training used frozen inputs from minit2i-frozen-inputs-v1


![[Krea2 Turbo] Asian Mix Lora](/media/civitai--2746042-krea2-turbo-asian-mix-lora/cover-320.webp)
![[Krea2 Turbo] Asian Mix Lora](/media/civitai--2746042-krea2-turbo-asian-mix-lora/cover.webp)




Comments