V2.0 UPDATE:
Pumped up training with 40 extra images, training resolution from 1024px to 1280px, network size from 16 to 64, 5000 steps.
I think the quality of the output has improved drastically!!!
V1.0:
I trained this lora for Flux locally in ComfyUI using a dataset of 100 images, all verbally described (thanks to Joy caption pre alpha) and trained for 2480 steps.
In the prompt you have to describe as many details as you can.
Use the keywords Westeros, Game of Thrones, Fantasy to activate the Lora.
About settings:
Fantastic creatures and locations work best with low cfg (2 to 3) - if the shot has some human characters it's better with mid-high cfg (3.5 to 7).
I think a lora weight of 0.8-0.85 gives good results.
You can find my workflow in the image without any watermarks or post-production.
Enjoy and please share here your creations!!!
描述:
训练词语: Westeros,Fantasy,Game of thrones
名称: Westeros_v3_rank16_bf16-step02485.safetensors
大小 (KB): 149818
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
