V4.5

回覆動慢風格

更鮮豔的顏色和細緻的臉孔

Slow and repetitive style

Brighter colours and more detailed faces.

V3_photoart

The actual effect is closer to 2.5D-3D game cg and its mergerd with SDXL photo-realistic relationship and the merging of the art department's big models, which reduces the recognition of the anime characters and brings better support for the real portraits and art-like LORA.

實際效果更接近2.5D-3D game cg

和SDXL 真實照片係還有藝術系的大模型合併,降低了對動漫人物的辨識度帶來對真實人像 藝術類LORA更好的支援

Sugest : WebUi re-Forge
Steps: 12

Sampler: LCM,

Schedule type: Automatic,

CFG scale: 1.5-2.5, (4 without FreeU, 1.5-2.5 with FreeU)

Model: UnstableInkDreamXL-V3_photoart,

Clip skip: 2,

Lora : negXL:0.5

V3.3

Stable output with art LORA.

Slightly eliminates the influence of anime and manga

Sugest : WebUi re-Forge
Steps: 12

Sampler: LCM,

Schedule type: Automatic,

CFG scale: 1.5-2.5, (2 without FreeU, 1.25-1.5 with FreeU)

Model: UnstableInkDreamXL-V,

Clip skip: 2,

Lora : negXL:0.5

V2.0

Minor Improvements

Add some artistic LORA adjustments

Sugest : WebUi re-Forge
Steps: 8-16, (10-16 without FreeU, 6-8 with FreeU)

Sampler: LCM,

Schedule type: Automatic,

CFG scale: 1.5-2.5, (2 without FreeU, 1.25-1.5 with FreeU)

Model: UnstableInkDreamXL-V02,

Clip skip: 2,

Lora : negXL:0.5

V1.0

My first test of illustration checkpoint merge

Old sd1.5 checkpoint still well for style paintings,But ....

https://civarchive.com/models/1540

Using WEBUI forge,with LCM sampling method and CFG 3~3.5,step 8-20

examples

Clip skip 2 

LCM/Autometic 

8-20 steps

CFG 3-3.5

640*896 resolutions or higher

prompt

masterpiece, best_quality,absurdres, highres,

negative

worst_quality, bad_quality, bad_hands,very_displeasing,lowres,extra_digit,fewer_ digits,jpeg ,watermark, instagram_username, twitter_username, patreon_username,frame,comic_fram, artifacts, signature, username, reference,