LoRA Training improvements

So I used an online metadata viewer to look at my Flux LoRAs. I have some suggestions.
Most importantly: "ss_fp8_base": "True" - seriously? This whole time I’ve been training in FP8? Please tell us if we’re training like this, I’d understand if it’s to cut server costs and wouldn’t mind at all, but not telling us isn’t too nice….
Other improvements:
1) Add the ability to set the following parameters [Note: I WOULD GLADLY PAY EXTRA FOR THESE, make them individual “add-on” options]:
"ss_face_crop_aug_range" - This one can be such a huge help, especially with datasets that have limited face close-ups if training a character
"ss_caption_dropout_rate" - Can lead to better training, worth adding as an option
"ss_apply_t5_attn_mask" - Allow us to turn this on for extra fee
"ss_timestep_sampling": "sigmoid" - Apparently changing it to “flux_shift” can result in dramatically better results based on an Issue in the official Kohya repo

2) Another “extra”: Allow training up to an image size of 1280×1280 (or more?) for an extra cost (which I realize would need to be hefty, and that’s fine)

3) METADATA REMOVAL: I was very surprised to find a variety of information about my Lora in the metadata viewer, including all my training settings and amount of images and repeats used. At NO COST, I feel you should allow people to have all this metadata removed. People steal it for their own purposes.

Thank you for your time, and all respect intended. Cheers.