Many models that I downloaded has very poor quality. Thats why I train myself, controlling all process. This model is much better than others.
Specially for portraits and closeup portraits. Selected EYES LOOK AT CAMERA (celebs shots usually look a side). HD and sharp images selected to dataset. Used 100-150 photos.
Trained on Comfyui FluxTrainer on 16gb Vram
If making big face, better use resolution 768 (minimize face-body proportion distortions) and than upscale whatever you want.
Small size of LORA because of training only 2 blocks: 7 and 20.
Demo images make with gguf 4bit, hyper8step and double pass.
More steps for training with AdamW is better. This is 7000 steps LORA
描述:
训练词语:
名称: hud_768_adamw_rank128_bf16-step07000.safetensors
大小 (KB): 9221
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
