This model is not trained from scratch. It is based off of SD 1.5's weights.
I'm sorry, but this is absolutely not a model trained from scratch. It was absolutely trained on-top of SD 1.5 and not just based on the architecture and not without the LAION dataset.
I compared it to the base SD 1.5 model from stability released by runway at the time, using https://huggingface.co/JosephusCheung/ASimilarityCalculatior and also made a ComfyUI node equivalent here: https://github.com/saftle/uber_comfy_nodes
When comparing with your model, it shows "Similarity: 97.59% (compared 11 blocks)" via my comfyUI node.
In comparison, if you compare base SD 1.5 with https://huggingface.co/aipicasso/clean-diffusion-2-0-poc, it shows up as a negative percentage because the weights are completely different as they should be. Clean Diffusion is actually a model trained from scratch.
Thank you for your feedback regarding our Free-Image-Generation-CC0 model. We sincerely apologize for any confusion caused by our previous description and naming.
We now understand that our original licensing and naming were misleading. While we were aware that SD 1.5 has its own license, we were uncertain which license to apply to our model because it generates images under CC0. This led us to label the model as CC0, even though the base weights (SD 1.5) are not CC0.
We would also like to clarify that all LoRA modules used in the merge are published under CC0-1.0 on Hugging Face, making them legally compatible with releasing the generated images under CC0.
To address this, we have updated the README to clearly explain that:
Generated images are CC0 and freely usable.
Base model weights are derived from SD 1.5 and are not CC0, while all LoRA modules are CC0.
The updated model and README can be found here: Free-Image-Generation
It was never our intention to mislead anyone regarding licensing, and we greatly appreciate your guidance in helping us maintain clarity, transparency, and integrity.