Meet Bivya!
This LoRA is trained on top of Pony Realism Checkpoint. Bivya was trained with 60 FLUX generated images and hand-crafted Tags. (If you want to learn more about LoRA Face training on Civit, just drop a comment and I'll pick it later).
Training Hints (Face/Character LoRA)
Data Prep is key - 15 images (10 similar, 5 diff angles) good to go, We went 60!
"Tags work better than captions" for LoRA training on SD and Pony.
How to tag for face training? For best results, It's split into static tags and dynamic tags!
Static Tags:
Name of character
Anatomy of face (detailed)
Hair color and length
Dynamic Tags:
Looking at viewer/ Looking away (Gaze)
Eye Level/ above angle/ below angle/ back angle/ over the shoulder (Camera Angle)
Expression (go detailed here, add multiple tags relevant to the image. eg: eye wink, big smile, curious expression, confident look etc)
Training PMs
"unetLR": 0.001,
"clipSkip": 1, [2 is for anime and cartoons]
"numRepeats": 7,
"resolution": 512,
"lrScheduler": "cosine",
"targetSteps": 1050,
"optimizerType": "Prodigy",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 20,
"shuffleCaption": false,
"trainBatchSize": 8
[Other PMs leave at default]Trigger & Strength
To include this LoRA in your generation use the main keyword, "Bivya" or "photo of Bivya".
Set the LoRA strength to 0.7-0.8
Generation best settings
CFG Scale: 6-8
Sampler: DPM++ 2M Karras (Euler doesn't give accurate results, but its pretty good too)
Steps: 30+
VAE: None (got better results than with detailer VAE)
Prompt Anatomy
Prompt: 1girl, solo, Bivya, <body_specs>, <pose>, <expression>, <location>, <lighting><shot/style_details>
Since the training set is mostly images of the face and sitting upright, you may have to use () for poses and mentioned the body pose before going for expressions.
Maintaining prompt anatomy order is crucial for better results
For better body control, use: Body Type Slider (It has some incredible cohesion with Pony Realism)
For pose, Don't use anime/cartoon styled pose LoRA, as it really messes up the overall realism and natural look, instead try to use vanilla pony realism to set up poses
Bivya's expressions while training were very minimal, if you push the LoRA to be more expressive you may not get good results.
For good lighting control, I recommend Pony Realism's v2.2 lighting guide
Negative: (Deformed eyes), Bad eyes, (Deformed Face), (Deformed Teeth), less detail, over expressive
These are some important negative that gave us the best execution results.
Add other negatives required for Pony Realism.
Tip: Using score_9, score_8_up, score_7_up & neg: score_6, score_5, score_4 has given us pretty bad results. From our understanding, these prompts generate images from a specific pool of characters (not characters, actually its good looking people) hence, messes with any face/character lora while generation.
Check out our example creations for a better understanding
Variations
Variations are the different expressions, camera angles and gaze, which we specifically trained to give users a lot of control for generating with Bivya. Some common variations of Bivya include,
Gaze: looking at viewer, looking away, looking to the side, looking over shoulder
Camera Angles: side angle, front angle, over the shoulder angle, top angle
Expressions:
gentle smile, bright smile, big smile with teeth, warm smile, subtle smile, playful smile, [most smiles might look the same, nonetheless we trained it xD]
intense gaze, sexy look, confident expression, curious expression, neutral expression, thoughtful expression, looking over shoulder, relaxed expression, elegant expression, caught off-guard, surprised expression, calm expression.
Other Objects: hands shielding the face, eye wink, looking upward, hand on forehead, lips slightly parted, hair braid, necklace, messy bun