Single file pure HTML tool for viewing LoRA metadata that can be run locally on your web browser without the need of installation or internet connectivity.
To use, just extract the attached Zip and open the extracted HTML file with the web browser of your choice. Alternatively, you can use the version hosted on Github: https://xypher7.github.io/lora-metadata-viewer
Not thoroughly tested, but should work with most modern web browsers. To be honest I didn't really intend to share this so its kind of low effort and pretty basic. However, I decided to post it since others might find it useful too. I don't really plan to update this as it already does what I need, plus I want to keep it as simple and lightweight as possible. Though suggestions are always welcome.
Github repository URL: https://github.com/Xypher7/lora-metadata-viewer
Features:
Configurable metadata summary
CivitAI resource lookup. (Requires internet connection)
Configurable training tag summary
Edit or remove metadata in the Safetensors file
Doro :3
Description
Added Arc En Ciel resource lookup option
Added web proxy settings for Arc En Ciel resource lookup
Added theme options
Added support for video preview
Added version update verification option
FAQ
Comments (5)
Doesn't work for Qwen loras.
It does, as long as they are in safetensors format. The problem is that Qwen loras typically don't include any useful metadata. However, this is mostly dependent on the creators and the tools they used to create the lora.
Can you please also add the "Average Training Speed" to the Summary?
In two formats "seconds pro step (s/it)" and "steps pro second (it/s)"?
Like what the kohya_ss displays during training.
(if you're making more than 1 iterations pro second it displays it/s value, if less, it switches to seconds pro iteration s/it value).
So basically for s/it, convert "Train Time" into seconds, then do "Train Time (sec) / Steps" => s/it. For it/s "Steps/Train Time (sec)" => it/s.
Example:
Train Time: 1h 13m 0s
Steps: 1170
Av. Training Speed: 3,74 s/it (0,27 it/s)
This will help to see how fast the training was on average.
Though it doesnt account for the time spent on sampling (if sampling was enabled during training) and etc., but its currently the best way i know of to calculate avr. speed, since its not saved into metadata.
weird question, but would it be possible for this to be able to read metadata from pixai or niji pictures? 🤔
mostly prompts, rather than models and all that jazz, tbh. specially for pixai... (also now i remembered that it do can read niji/mj metadata nvm so ig just pixai)