編輯嚴選
挑戰真相片+賽博風 - ControlNet三合一 | Stable Diffusion

2023/07/06閱讀時間約 12 分鐘


T2I
-融入真人姿態

因為想嘗試將真實照片中的人物姿態融入成品中,所以在Text2Image(文生圖)階段使用了三種ControlNet,包含非ControlNet原作者提供的光源模型:

Depth Map

為了保護隱私,只提供ControlNet運算出來的真實照片Depth map:
Depth Map | ControlNet
ControlNet的設定如下圖:
Depth settings | ControlNet

reference_only

reference_only的參考圖如下:
reference_only | ControlNet
註1:我在網路上蒐不到這張的原作者是誰,唯一可能有關的推特帳號請按這裡前往.
註2:如果知道原作者或原作者有看到本篇,煩請告知我好附上來源。If the original author or anyone knowing the author sees this, please kindly notify me to add credit.
ControlNet的設定如下圖:
reference_only settings | ControlNet

打光圖

打光圖我很粗略畫了一張:
light map | ControlNet
ControlNet的設定如下圖:
lightingBasedPicture settings | ControlNet

其他設定(整段複製貼到T2I的positive prompts即可套用):

a female adult cyborg and a female child android waiting for green light on the sidewalk at night, (detailed faces), (extremely detailed), heavy rain, futurisitic, magic and technology, masterpiece, abs res, best quality, sci-fi scene, dark environment, dystopia, cityscape, downtown, cyberpunk, water puddles, water splashes, rain drops, Tron, bodysuit, prosthetic legs, prosthetic arms, umbrella, mechnical parts, mechnical equipments, tools, machine components, robots, spaceships, ACG, Japanese anime, (from behind),
Negative prompt: bad-hands-5, ng_deepnegative_v1_75t, extra fingers, deformed hands, polydactyl, ((low quality, worst quality, monochrome, greyscale, grayscale, watermark, text, blurry, jpeg artifacts)), cropped, normal quality, ((signature, username, artist name, logo)), 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, skin spots, acnes, skin blemishes, age spot,
Steps: 25, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 4009561042, Size: 1104x1680, Model hash: 4199bcdd14, Model: revAnimated_v122, VAE: vae-ft-mse-840000-ema-pruned, Denoising strength: 0.5, Clip skip: 2, Version: 875d0db, Parser: Full parser, ControlNet 0: "preprocessor: tile_resample, model: control_v11f1e_sd15_tile [a371b31b], weight: 1, starting/ending: (0, 1), resize mode: Just Resize, pixel perfect: True, control mode: Balanced, preprocessor params: (64, 1, 64)"
T2I、I2I的VAE都是vae-ft-mse-840000-ema-pruned,可由此下載。有.ckpt和.safetensor兩種版本,使用其一即可。

T2I階段保留了四張圖:

1100036977
1100036978
1100036979
1100036980

I2I-放大/增添細節

tile_resample

I2I只用了tile_resample這個很好用的ControlNet模型:
tile_resample settings | ControlNet
*tile_resample和ControlNet的基本用法可參考筆者這篇: 忠實放大、增添細節 ControlNet 1.1 - Tile Resample - Stable Diffusion

其他設定(整段複製貼到I2I的positive prompt即可套用)

a female adult cyborg and a female child android waiting for green light on the sidewalk at night, (detailed faces), (extremely detailed), heavy rain, futurisitic, magic and technology, masterpiece, abs res, best quality, sci-fi scene, dark environment, dystopia, cityscape, downtown, cyberpunk, water puddles, water splashes, rain drops, Tron, bodysuit, prosthetic legs, prosthetic arms, umbrella, mechnical parts, mechnical equipments, tools, machine components, robots, spaceships, ACG, Japanese anime, (from behind),
Negative prompt: bad-hands-5, ng_deepnegative_v1_75t, extra fingers, deformed hands, polydactyl, ((low quality, worst quality, monochrome, greyscale, grayscale, watermark, text, blurry, jpeg artifacts)), cropped, normal quality, ((signature, username, artist name, logo)), 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, skin spots, acnes, skin blemishes, age spot,
Steps: 25, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 4009561042, Size: 1104x1680, Model hash: 4199bcdd14, Model: revAnimated_v122, VAE: vae-ft-mse-840000-ema-pruned, Denoising strength: 0.5, Clip skip: 2, Version: 875d0db, Parser: Full parser, ControlNet 0: "preprocessor: tile_resample, model: control_v11f1e_sd15_tile [a371b31b], weight: 1, starting/ending: (0, 1), resize mode: Just Resize, pixel perfect: True, control mode: Balanced, preprocessor params: (64, 1, 64)"

從T2I選了兩張衣著有發光的圖片以I2I放大和增添細節:

4009561042
2657899963

心得-打光仍在起步,tile_resample仍很好用

  1. ControlNet的打光模組現階段仍不是很好用,並且只能用在T2I;
    另一方面,能控制光的強弱分布是很強大的功能,期待未來也能用在I2I上。
  2. 對有些checkpoint模組而言,好比這次嘗試使用的SweetMix、ReV Animated,extremely detailed、detailed face等強調細節的提示詞有很顯著的影響。
  3. ControlNet的tile_resample在I2I放大圖片時還是很好用,能大幅降低在 Denoising Strength > 0.4 時冒出莫名其妙的物件,同時又保留了相當程度的變化。
祝大家算圖愉快!
為什麼會看到廣告
78會員
104內容數
筆者探索AI工具時遇到的難題、解方和雜談,不定期更新。
留言0
查看全部
發表第一個留言支持創作者!