I recently trained my first LoRa and this tool was very useful for organizing the images:
For prompting the descriptive text I've been using JoyCaption:
Demo: https://huggingface.co/spaces/fancyfeast/joy-caption-pre-alpha
//---//
The batchcropper tool linked at the top allows you to paste the Joycaption prompts alongside the images to create your own dataset
//---//
From the SD3.5-M model release we get this information regarding the batch size for the T5 model:
Link: https://huggingface.co/stabilityai/stable-diffusion-3-medium
The T5 is an LLM model. Input can be very different from the CLIP models many are used to prompt with.
How can this be used?
One thing I've been experimenting with is taking audio segments from various places like the Charcarhadron 40K lore video (good listen to btw) : https://youtube.com/@adeptus-astra?si=zUl7t8wIEromruL2
Downloading it as an MP3 either via an online tool or if the video is long (more than 1h , or if you want to download many videos to mp3 at once) using this notebook I've coded: https://huggingface.co/codeShare/JupyterNotebooks/blob/main/YT-playlist-to-mp3.ipynb
And then passing the MP3 through this online transcriber tool: https://turboscribe.ai
And voila! Now you have bits and pieces of rather unique prompt snippets which can be used for the T5.
I made some spaceships using this technique:
See this post for prompt: https://tensor.art/images/790655766681080995?post_id=790655766676886694&source_id=njq1pFzjlEOwpPEpaXny-xcu
LoRa which I trained on the SD3.5M model: https://tensor.art/models/790774208985724051?source_id=njq1pFzjlEOwpPEpaXny-xcu
I used the training feature on TensorArt. Its the Symbol with barbell called "Training" which can be found by clicking the button on the left side ---> of your TensorArt UI
Cheers,
Adcom
