NVIDIA驅動程式最新版本532.03
因為NVIDIA驅動程式最新版本532.03據說在Windows 11環境下能大幅提升Stable Diffusion的inference效能(簡單的說即是算圖效能),所以我更新了驅動程式後想測試看看。
先說結論:好像有,又好像沒有,總之希望有。
因為只是測試,所以這次的正面提示詞寫得很簡單,且剛好又想試試看能否畫出《天堂II》的黑暗精靈,於是+/-提示詞如下:
+
master piece, dark elf, lineage 2, female, dancer, grey skin, gray skin, blue skin, pale skin,
-
bad-hands-5, ng_deepnegative_v1_75t, (extra fingers, deformed hands, polydactyl:1.4), (low quality, worst quality:1.4), (monochrome:1.1), (greyscale), watermark, text, blurry, jpeg artifacts, cropped, normal quality, signature, username, artist name, cartoon, canvas frame, lowres, (disfigured), (bad art), (deformed), (extra limbs), (b&w), weird colors, (duplicate), (morbid), (mutilated), mutated hands, (poorly drawn hands), (poorly drawn face), (mutation), (ugly), (bad proportions), cloned face, out of frame, gross proportions, (malformed limbs), (missing arms), (missing legs), (extra arms), (extra legs), fused fingers, (long neck), lowres, (grayscale), (skin spots), acnes, skin blemishes, (age spot),
成果如下:
黑暗精靈(?)大圖
這次測試的心得
個人覺得正面提示詞比較簡單時,ControlNet的tile_resample效果更佳。另一方面,Stable Diffusion不懂得點到為止的老毛病依然存在。
以下是幾個問題處和手動修正後的對照:
1. 衣服材質貼到皮膚上
2. 嘴唇紋路太過明顯
3. 添加過多的瀏海髮絲
4. 頭髮紋路貼到長手套上
這次測試是把512 x 768 的圖經過二次3倍放大,所以修正方式是把一次放大的圖局部覆蓋至二次放大的圖,不過嘴唇例外,是直接在繪圖軟體中後製。
提示詞「lineage 2」竟然有用!
這次測試最出乎我預料的是「lineage 2」(天堂2的英文)提示詞的效果非常明顯,而且無論是ACG風的SweetMix還是寫實風的FantasticMix都有效果。所有設定相同,差別只在有無「lineage 2」的對照圖如下:
搞怪的是,「lineage 2」加上「dark elf」(黑暗精靈的英文)提示詞,能讓衣著、五官、體型、髮色、身材等都更趨近《天堂2》的黑暗精靈,但偏偏最重要的膚色畫不太出來,並且不使用「lineage 2」提示詞時反而更能讓「grey skin, gray skin, blue skin, pale skin,」等提示詞發揮作用,讓膚色更趨近於黑暗精靈(只有SweetMix可以,FantasticMix則是畫出很搞笑的結果)。
整體而言還是很滿意。
補充
以下是產生512 x 768小圖時的設定,如果直接貼在Vlad Diffusion的正面提示詞中然後按提示詞框右邊的箭頭可直接套用。A1111的話可能要自行手動套用。
master piece, dark elf, lineage 2, female, dancer, grey skin, gray skin, blue skin, pale skin,
Negative prompt: bad-hands-5, ng_deepnegative_v1_75t, (extra fingers, deformed hands, polydactyl:1.4), (low quality, worst quality:1.4), (monochrome:1.1), (greyscale), watermark, text, blurry, jpeg artifacts, cropped, normal quality, signature, username, artist name, cartoon, canvas frame, lowres, (disfigured), (bad art), (deformed), (extra limbs), (b&w), weird colors, (duplicate), (morbid), (mutilated), mutated hands, (poorly drawn hands), (poorly drawn face), (mutation), (ugly), (bad proportions), cloned face, out of frame, gross proportions, (malformed limbs), (missing arms), (missing legs), (extra arms), (extra legs), fused fingers, (long neck), lowres, (grayscale), (skin spots), acnes, skin blemishes, (age spot),
Steps: 20, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 704885009, Size: 512x768, Model hash: 76385703e8, Model: sweetMix_v15, Clip skip: 2, Token merging ratio: 0.3, Token merging ratio hr: 0.3
雜談:Stable Diffusion - XL
最近stability ai正在測試下世代的
Stable Diffusion - XL,據稱可使用更直覺式的簡單提示詞算出更好的結果,且
手腳部位的正確度大幅提升(這才是重點!)。現在可在
ClipDrop的網站試玩,但每生一張圖就要等很久才能再升一次張。
雖然我對使用這次測試的正面提示詞產生的結果不甚滿意:
據說XL骨子裡更像Midjourney,所以才是SD-XL而不是SD-2。但如果目前既存的SD模組和插件都不支援的話,我們這些玩家可就頭大了呢。
—stability ai不如想辦法先把目前的手腳問題搞定?
祝大家算圖愉快!