SSD 1B Anime

🚀 SSD-1B Anime Model

This is a text-to-image model based on Stable Diffusion, created through a specific merging and distillation process.

🚀 Quick Start

The safetensors file can be used in the latest ComfyUI.

✨ Features

Model Creation Steps

This model was created in the following two steps:

1. Merge SSD-1B with 1.3 times the difference between NekorayXL and sdxl-1.0. For the keys before and after distillation, this mapping is assumed.
2. Distillation using the difference from the final output of NekorayXL as the loss (learning rate 1e-5, batch size 4, 23000 steps).

LoRA

ssd-1b-anime-cfgdistill: This is a LoRA trained to generate decent images at cfg_scale = 1. When cfg_scale = 1, the calculation on the negative side is not required, so the computational cost is halved. If it is greater than 1, you won't get the benefit of computational cost reduction, but it seems to work as a normal performance-improving LoRA. However, it is recommended to use a lower value than normal generation.

LCM

lcm-ssd1b-anime: This is trained from SSD-1B's LCM-LoRA.

Key Correspondence between SSD-1B and SDXL

Since it's not clear which modules were deleted, we estimated using cosine similarity. Since only the Transformer_depth has changed (presumably), we investigated the parameters of the Attention layer in SDXL and SSD-1B. When it changes from 2 layers to 1 layer, the first layer remains. When it changes from 10 layers to 4 layers, the 1st, 2nd, 3rd, and 7th layers remain.

※ The third layer in the up layer remains 10 layers, but the cosine similarity result was strange. For now, we assume this part hasn't changed.