Z-image is pretty good.Basically, it can run with or without trigger words. My dataset tags include trigger words, but I feel it can still achieve good results without them. Using natural language prompts would be better, but word prompts are also acceptable.
The prompts can be generated using the system prompts from the official code and use LLM function,the results are quite good.
Switched up the dataset this time. The old tags just weren't cutting it for the image features, so I fixed those and used a new model for training.other settings are basically unchanged, almost the same as v1.0
Sampler: Euler A
Schedule type: Simple
Steps: 30
CFG: 6~7
Hires. Fix: R-ESRGAN 4x + Anime6B
Magnification: 2
Prompt Prefix:
at0ka, 1girl, solo, school uniform, hat, long sleeves, brown eyes, boots, belt, knee boots, luna nova school uniformNegative Prompt:
????, worst quality, old, early, low quality, lowres, signature, username, logo, (((bad hands))), mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro,bad legs,multiple legs,multiplayer,Sampler: Euler
Schedule type: Simple
Steps: 30
CFG: 6~7
Hires. Fix: R-ESRGAN 4x + Anime6B
Magnification: 2
Prompt Prefix:
at0ka, 1girl, solo, long hair, school uniform, hat, long sleeves, brown eyes, boots, belt, knee boots, shirt, luna nova school uniformNegative Prompt:
????, worst quality, old, early, low quality, lowres, signature, username, logo, (((bad hands))), mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro,bad legs,multiple legs,multiplayer,
Overall, it's good，looking forward to the base. This time, there was some underfitting because the dataset is quite large, and the step size was set too small. A larger step size should have resulted in a better fit.
1. The rights to reposted models belong to original creators.
2. Original creators should contact SeaArt.AI staff through official channels to claim their models. We are committed to protecting every creator's rights. Click to Claim
