This time I found that ZIT training actually resolution settings can choose 512/256 "resolution": [512, 256], "lr": 0.00015, as well, save VRAM also save time, fitting is also super fast
In fact, 750 steps can see 2000 steps is affected by the training set too cartoony, so if you want realistic results please reduce the weight to 0.3 below, and add, COSPLAY photos, present character appearance, hair color, eye color, makeup drama, costume and prop details prominent. The pose is close to the original stand-up, the background matches the world view. The light strengthens the material and eye details, improving contrast and special effects in post-production, making the overall look like a character poster rather than an everyday portrait. COSPLAY photo，。
ZIT版本感覺有點做過頭了
這次發現ZIT訓練時其實resolution設定可以選擇512/256"resolution": [512, 256],"lr": 0.00015,就好,省VRAM也省時間,擬合也超快
其實750步就能看2000步就被訓練集影響到太卡通化了,所以如果想要擬真結果請把權重降低到0.3以下,並且加上,COSPLAY照片，呈現角色化外觀，髮色、瞳色、妝容戲劇化，服裝與道具細節突出。姿勢接近原作立繪，背景配合世界觀。光線強化材質與眼部細節，後期提升對比與特效，整體像角色海報而非日常人像。 COSPLAY photo,
https://en.wikipedia.org/wiki/Powered_exoskeleton
V1 trained by photo
V2 = V1 +non-humanoid robot lora +.... to get model more anime scifi like.....
If use full body cause of the small face ratio, it is recommended to use ADTAILER or HIRES.FIX to avoid face crushing.
but if you use prompt close-up............more like cyborg than exoskeleton
and can mix my other mecha robot lora