A pathetic attempt at fixing hand, first image preview is On, second Off.
commun trigger words: a person's hand
Datasets:
11K Hands by Afifi, Mahmoud, Cleaned blurry and similar using script, 2 down to 3158 Images, available for download in the Training Data.
Training settings:
"0.0002:10, 0.00001:20, 0.0007:60, 0.00005:200, 0.0002:500, 0.0001:3000, 0.00001" for 37K steps, and then 81K with 0.001+Clip_grad: value.
Description
LoRA couldn't learn as many images; if split and retained, it experiences some catastrophic forgetting. Therefore, I opted for textual inversion.
Feel free to train my dataset on LoRA and publish it without crediting me 🎩
FAQ
Comments (17)
how to use it?
It's a TI, just put it in your embeddings folder then call the trigger word in your prompt
On the negatives: Fixhand
Insert is not a panacea BUT greatly improves fingers and reduces curvature. HOWEVER, art will change a bit. It is better to use it as a fix, it works on many lora.
just put Fixhand in the negative is fine?
@a56622Â See my examples. ALL of them compare without Fixhand and Fixhand in the negatives. Try inserting Fixhand into simple prompts 1girl, in full growth, best quality, masterpiece, good quality,Fixhand, and then with them, new negatives, the result will surprise you.
what is data training for?
where should I put training data
I've shared my Training Data for anyone who wants to train. If you are not planning to fine-tune a model or LoRa, feel free to ignore it.
hello, can you provide the safetensors?
Added, script used.
this embedding is to put in negative or prompt
Prompt
This one throws errors in the Civitai generator.
I've had very good results with this embedd. I made like 400 pictures of which i had to inpaint less than 10, and it got it fixed on first try. Thank you.
So is it for positive or negative prompt?
Positive, see metadata prompt of sample preview.
DigitalDreamer thxx
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.