This model is of Amanda Tapping in her role as Samantha Carter, it is part of a series of 4 models with which you can recreate the complete SG-1 team. It is my first series of LoRAs.
Sam is trained to appear in the SG1 uniform.
The trigger word is "sgsam", but it is not necessary as it will turn any woman into Sam.
Links to all the SG-1 members and other Stargate characters
JACK O'NEILL https://civarchive.com/models/13792/jack-oneill-stargate
SAMANTHA CARTER https://civarchive.com/models/13748/samantha-carter-stargate
TEAL'C https://civarchive.com/models/13822/tealc-stargate
DANIEL JACKSON https://civarchive.com/models/13818/daniel-jackson-stargate
GENERAL HAMMOND https://civarchive.com/models/17165/major-general-hammond-stargate
ASGARD RACE https://civarchive.com/models/15273/asgard-race-stargate
Description
FAQ
Comments (13)
hah, what a timing with this one :) XD
OMG I just saw your Sam model, I'm glad to see more stargate fans over here :)
This works really nicely, great job, do you mind saying how many photos and how many steps you used to train the LORA? I'm just starting and getting results that are close but not recognizable (different person)
Thank you! I'm glad you like it.
Basically, what I did was follow the Aitrepreneur tutorial (https://youtu.be/70H03cv57-o). Once I had trained the first Lora, I looked for its weaknesses and trained another Lora to solve those weaknesses. For example, if the first Lora had difficulty creating full-body images, the second Lora would have many more full-body images. After that, I merged the two Loras using the kohya merge Loras function and compared the results between Lora 1, Lora 2, and the merged Lora.
For Carter, I made a total of three Loras: Lora 1 (31 images) and Lora 2 (29 images) were mixed at 60%-40%, and the result was mixed with Lora 3 (26 images) at 50%-50%.
One thing I highly recommend is testing the Lora on different models, as some models will give you more accurate results than others for your character. For example, in my experience, RealVision creates highly realistic images but doesn't resemble my characters as much. However, others like Protgen or Deliberate are more faithful to my characters.
I hope this helps you.
@averok When you say "testing on different models" do you mean using different models when training? Or just when using the LORA? I am training with a similar number of images (I started with 50, now am using 100) What are you using for the steps when training? I had some people suggest I set it to 25 or 50, not great results so I'm trying 100 now. Right now I'm training with clarity, if that doesn't work I'll try deliberate. Though some people suggested I just use SD1.5
@clevenger I mean using the LORA, for training I have used SD1.5 and 100 steps.
@averok This was REALLY helpful. Thank you.
great LoRA, thanks for the addition...
Awesome, thank you so much!!!
Thank you! I'm glad you like it!
nice work! question: did you create your own ControlNet pose for your samples or did you use a standard one? would you mind sharing the pose you used for the full body with headshot in foreground? again, nice work and thanks!
Thank you!! I created my pose for ControlNet. I just uploaded it, here is the link! https://civitai.com/models/99493/thumbnail-pose
@averok awesome - thanks so much!