Just now, we have preliminarily completed the establishment of the v1.5 version of the LoRA automatic training pipeline, with the entire process successfully executed. Based on the initial test results, the model has shown significant quality improvement in almost all aspects.
The main updates of this version (not all, a more detailed article will follow) include:
Building a larger dataset based on a comprehensive web search, with many waifu datasets exceeding 1,000 images.
Increased training resolution, from the original 512x512 to 720x720.
A comprehensive automatic step selection algorithm covering aspects such as character feature fidelity, image style quality, and prompt controllability.
Based on the step selection, we can now stably automate training for small datasets (less than 50 original images).
Here are a few models that have been released for you to try, and we welcome any suggestions. We will also make some parameter adjustments according to the actual situation to ensure that the automatically trained models in the future have higher quality:
Web-based large dataset training example (about 1,500 images): https://civitai.com/models/121400?modelVersionId=314949
Anime-based large dataset training example (over 2,000 images): https://civitai.com/models/149687?modelVersionId=314968
Web-based medium-sized dataset training example (less than 1,000 images): https://civitai.com/models/279670/moriyama-shiemi
Web-based small dataset training example (about 150 images): https://civitai.com/models/279675/kristen-wright-arknights
Web-based tiny dataset training example (about 40 images): https://civitai.com/models/279681/leto-arknights
Once the new version of the pipeline enters a stable state, we will publish a more comprehensive article introducing the new methods and technologies used in the new version of the pipeline, as well as a brief introduction to some mature open-source tools, to help more people.
So, stay tuned!
Update1:
A bad news is, bundled lora (the format my LoRAs are using, it is now officially supported by a1111's webui without any extra plugins required, see: https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/13568) are not supported by civitai for now.
So the online image generation service cannot draw the waifu. Obviously the trigger words are not triggered properly. Like this
Maybe I have to open a ticket to civitai.
Update 2:
Due to the pivotal tuned LoRA will not supported by civitai onsite generator in a short time. So we decide to turn off the onsite generator of the new pulished models.
Actually we are willing to see our models to be used in a convenient way like this. By unfortunately it doesn't work on it. We will turn it on again when civitai support the generation of pivotal tuned LoRAs.
Update 3:
Model of shu (a new dragon waifu from the newest arknights CN server PV) published: https://civitai.com/models/283562/shu-arknights
About 40 medium-quality images was used (cropped to 113).
The only things I do to create it were:
Get her dataset, like this. And then here dataset will be uploaded to huggingface 15 mins later: https://huggingface.co/datasets/CyberHarem/shu_arknights
Train her and publish her model with CyberHarem project (README still WIP for now): https://huggingface.co/CyberHarem/shu_arknights . Meanwhile, I'm playing my palworld for hours.
Publish here mode to civitai like this
That's all. There is nothing else I need to do, lol.
Update 4:
A good news: v1.5 works well on small samples (<50 original images) web-based loras. This has been proved on 100+ loras.
A bad news: not work well on small samples anime-based loras. the image quality or something will extremely seriously affect the quality of lora. so, maybe it's time for solving the quality problem of anime videos.