image-2模型在生成二次元图片时貌似会出现问题
- 内容介绍
- 文章标签
- 相关推荐
如题,这几天用image-2模型疯狂生成角色插图的时候,每次遇到“室外““居家“这种场景,生成出来的图片,明暗色块的排布都会变的像马赛克一样很规律的排布。
不知道各位佬友遇到过没有
IMG05131828×1006 416 KB
--【壹】--:
是这样子的,我要求二次元画风的时候经常出现这个像马赛克一样的颗粒感,不清楚怎么回事
--【贰】--:
我在labnana的发现里面,看见别人生成的照片也会有一样的问题,应该就是模型的问题了
image700×688 172 KB
--【叁】--:
1000027253.png2160×3040 1.42 MB
1000027254.png2160×3040 1.42 MB
满血没有问题
--【肆】--:
除了API,部分第三方网站也有,比如 Labnana现在每人有免费的100张gpt-image-2额度试用 提到的、Genspark(免费账户可以开4K+中等质量,虽然数量少)等,你可以先用相同的提示词试试看
--【伍】--:
最好的解决办法是给参考图垫图,纯文生图就在提示词限定画风,或者加上负面提示词试一试【降噪、免锐化、禁图章复制感】禁满屏白色高光噪点,禁止对画面进行锐化处理。
--【陆】--:
这样吗,谢谢佬,我晚上把你的负面提示词加上再试试
--【柒】--:
给个提示词佬,谢谢了,自己写的实在是太笨了
--【捌】--:
image1672×941 412 KB
提示词
CORE PROMPT(核心语义锚点) Cinematic anime poster featuring a confident man and a sharp-witted woman facing each other across a casino table, intense eye contact, psychological duel atmosphere, playing cards floating mid-air, high-end anime illustration style, dramatic tension, modern nightlife setting, immersive storytelling STYLE(风格层) High-end anime key visual, Japanese commercial illustration style, ultra-clean lineart with refined cel-shading, soft gradient rendering, glossy highlights, cinematic depth, warm golden lighting palette, polished studio-quality finish, modern anime promotional artwork CHARACTER DESIGN(人物设定层) Male character: handsome young man with messy black hair, confident smirk, relaxed posture, wearing a slightly unbuttoned white shirt, modern stylish look Female character: elegant young woman with long flowing orange hair, sharp gaze, composed and competitive expression, wearing a sleek black outfit with accessories, modern nightlife fashion, strong and intelligent presence FACE(面部控制层) Both characters with expressive anime-style faces, sharp eyes with vibrant highlights, subtle smirk on male, focused and slightly tense expression on female, emotional tension conveyed through eye contact, cinematic facial shading POSE(姿态叙事层) Leaning forward across the table, elbows resting on surface, close interpersonal distance, strong face-to-face composition, subtle power balance, woman holding playing card near lips, man casually engaging, dynamic conversational tension HANDS & PROPS(手部与道具层) Playing cards (spade suit visible), poker chips scattered on table, reflective glossy surface, subtle motion of cards in mid-air, hands positioned naturally with expressive gestures, casino game context clearly established COSTUME(服装层) Modern nightlife fashion styling, male in casual elegant white shirt with soft fabric folds, female in sleek black outfit with clean silhouette and minimal accessories (choker, bracelet, earrings), refined material rendering with subtle highlights HAIR(头发层) Male: messy layered black hair with soft volume and directional flow Female: long orange hair with smooth gradient highlights, natural flow, slightly lifted strands for motion, glossy anime hair rendering ENVIRONMENT(环境层) Casino or upscale bar interior, warm ambient lighting, blurred background with bokeh lights, depth of field effect, lively yet soft-focus crowd suggestion, cinematic nightlife atmosphere COMPOSITION(构图层) Horizontal cinematic framing (16:9), medium close-up shot, two-character symmetrical tension composition, table as foreground anchor, strong eye-line connection, shallow depth of field, subject isolation with soft background blur LIGHTING(灯光层) Warm golden key lighting, soft ambient fill light, subtle rim lighting on hair and shoulders, reflective highlights on table surface, cinematic contrast without harsh shadows, intimate indoor lighting mood COLOR SYSTEM(色彩系统) Dominant tones: warm gold, amber, black Accent colors: orange (hair), white (shirt), red/blue (chips) Balanced warm palette with vibrant highlights MOOD(情绪层) Tense, playful, competitive, flirtatious undertone (safe level), psychological game, confident energy, stylish and modern, engaging interpersonal chemistry Negative prompt(负面约束) low quality, blurry, bad anatomy, extra limbs, distorted hands, messy composition, dull colors, overexposed lighting, flat shading, 3D render look, text, watermark, logo,降噪、免锐化、禁图章复制感,禁满屏白色高光噪点,禁止对画面进行锐化处理 Aspect ratio(画幅) 16:9, cinematic widescreen, anime poster style, high-resolution, clean composition
这张是网页直出的,提示词隔壁生图大赛上一位佬的提示词,这张我感觉白色噪点,鳞片式上色都要少一点
--【玖】--: gpt-image-2生成宽图时有点奇怪的毛病 搞七捻三
openai预训练的渲染问题,现在image2生成的很多图片也是,有很多的噪声噪点gpt-image-2生成宽图时有点奇怪的毛病 搞七捻三
恰巧这就是image2的一个特点,也就是“碎玻璃”感,有些图有,有时候没有,但是,如果要求在第一张图的基础上生成类似的相似图,就会稳定复现
--【拾】--:
刚刚又试了一张加上负面提示词效果也不好
image1303×827 317 KB
还是非常明显,推测是风格相关提示词的问题
【High-end anime key visual, Japanese commercial illustration style, ultra-clean lineart with refined cel-shading, soft gradient rendering, glossy highlights, cinematic depth, warm golden lighting palette, polished studio-quality finish, modern anime promotional artwork】,这句比较关键
--【拾壹】--:
我在抖音上也看到有人说会有破碎拼图感,可能是一些新策略导致的副作用吧
--【拾贰】--:
我就是给的参考图,第一张出来一般都没问题,让他改一两次就又出现这种效果了
--【拾叁】--:
那现在就不只宽图了,4:3也会出现问题,而且细节越多的图,出现问题的机率就越大
--【拾肆】--:
满血版只能调用api吗使用吗,我现在直接是在网页端用的
--【拾伍】--:
看起来好像不只是二次元图片会有这个问题,之前看别人发的一些偏现实的也有这个问题,出来的图就好像 Stable Diffusion 采样器和步数设置的有问题一样
--【拾陆】--:
都这样 尤其是前几天到处发的海报,这种破碎纹理感几乎遍布图片,看起来一直觉得怪异
一直也没见什么人提
--【拾柒】--:
体感似乎是提示词和参考图的问题
没有风格指定就默认上碎玻璃风格
有风格指定就把这个覆盖掉了
--【拾捌】--:
很常见,有种马赛克模糊的感觉。我看到别人出的图也类似
--【拾玖】--:
以前的gpt image 也有类似问题,这次特别多。非二次元风格也遇到很多次,感觉是调用了量化版的模型。
有时候不得不去香蕉里洗一下图。
如题,这几天用image-2模型疯狂生成角色插图的时候,每次遇到“室外““居家“这种场景,生成出来的图片,明暗色块的排布都会变的像马赛克一样很规律的排布。
不知道各位佬友遇到过没有
IMG05131828×1006 416 KB
--【壹】--:
是这样子的,我要求二次元画风的时候经常出现这个像马赛克一样的颗粒感,不清楚怎么回事
--【贰】--:
我在labnana的发现里面,看见别人生成的照片也会有一样的问题,应该就是模型的问题了
image700×688 172 KB
--【叁】--:
1000027253.png2160×3040 1.42 MB
1000027254.png2160×3040 1.42 MB
满血没有问题
--【肆】--:
除了API,部分第三方网站也有,比如 Labnana现在每人有免费的100张gpt-image-2额度试用 提到的、Genspark(免费账户可以开4K+中等质量,虽然数量少)等,你可以先用相同的提示词试试看
--【伍】--:
最好的解决办法是给参考图垫图,纯文生图就在提示词限定画风,或者加上负面提示词试一试【降噪、免锐化、禁图章复制感】禁满屏白色高光噪点,禁止对画面进行锐化处理。
--【陆】--:
这样吗,谢谢佬,我晚上把你的负面提示词加上再试试
--【柒】--:
给个提示词佬,谢谢了,自己写的实在是太笨了
--【捌】--:
image1672×941 412 KB
提示词
CORE PROMPT(核心语义锚点) Cinematic anime poster featuring a confident man and a sharp-witted woman facing each other across a casino table, intense eye contact, psychological duel atmosphere, playing cards floating mid-air, high-end anime illustration style, dramatic tension, modern nightlife setting, immersive storytelling STYLE(风格层) High-end anime key visual, Japanese commercial illustration style, ultra-clean lineart with refined cel-shading, soft gradient rendering, glossy highlights, cinematic depth, warm golden lighting palette, polished studio-quality finish, modern anime promotional artwork CHARACTER DESIGN(人物设定层) Male character: handsome young man with messy black hair, confident smirk, relaxed posture, wearing a slightly unbuttoned white shirt, modern stylish look Female character: elegant young woman with long flowing orange hair, sharp gaze, composed and competitive expression, wearing a sleek black outfit with accessories, modern nightlife fashion, strong and intelligent presence FACE(面部控制层) Both characters with expressive anime-style faces, sharp eyes with vibrant highlights, subtle smirk on male, focused and slightly tense expression on female, emotional tension conveyed through eye contact, cinematic facial shading POSE(姿态叙事层) Leaning forward across the table, elbows resting on surface, close interpersonal distance, strong face-to-face composition, subtle power balance, woman holding playing card near lips, man casually engaging, dynamic conversational tension HANDS & PROPS(手部与道具层) Playing cards (spade suit visible), poker chips scattered on table, reflective glossy surface, subtle motion of cards in mid-air, hands positioned naturally with expressive gestures, casino game context clearly established COSTUME(服装层) Modern nightlife fashion styling, male in casual elegant white shirt with soft fabric folds, female in sleek black outfit with clean silhouette and minimal accessories (choker, bracelet, earrings), refined material rendering with subtle highlights HAIR(头发层) Male: messy layered black hair with soft volume and directional flow Female: long orange hair with smooth gradient highlights, natural flow, slightly lifted strands for motion, glossy anime hair rendering ENVIRONMENT(环境层) Casino or upscale bar interior, warm ambient lighting, blurred background with bokeh lights, depth of field effect, lively yet soft-focus crowd suggestion, cinematic nightlife atmosphere COMPOSITION(构图层) Horizontal cinematic framing (16:9), medium close-up shot, two-character symmetrical tension composition, table as foreground anchor, strong eye-line connection, shallow depth of field, subject isolation with soft background blur LIGHTING(灯光层) Warm golden key lighting, soft ambient fill light, subtle rim lighting on hair and shoulders, reflective highlights on table surface, cinematic contrast without harsh shadows, intimate indoor lighting mood COLOR SYSTEM(色彩系统) Dominant tones: warm gold, amber, black Accent colors: orange (hair), white (shirt), red/blue (chips) Balanced warm palette with vibrant highlights MOOD(情绪层) Tense, playful, competitive, flirtatious undertone (safe level), psychological game, confident energy, stylish and modern, engaging interpersonal chemistry Negative prompt(负面约束) low quality, blurry, bad anatomy, extra limbs, distorted hands, messy composition, dull colors, overexposed lighting, flat shading, 3D render look, text, watermark, logo,降噪、免锐化、禁图章复制感,禁满屏白色高光噪点,禁止对画面进行锐化处理 Aspect ratio(画幅) 16:9, cinematic widescreen, anime poster style, high-resolution, clean composition
这张是网页直出的,提示词隔壁生图大赛上一位佬的提示词,这张我感觉白色噪点,鳞片式上色都要少一点
--【玖】--: gpt-image-2生成宽图时有点奇怪的毛病 搞七捻三
openai预训练的渲染问题,现在image2生成的很多图片也是,有很多的噪声噪点gpt-image-2生成宽图时有点奇怪的毛病 搞七捻三
恰巧这就是image2的一个特点,也就是“碎玻璃”感,有些图有,有时候没有,但是,如果要求在第一张图的基础上生成类似的相似图,就会稳定复现
--【拾】--:
刚刚又试了一张加上负面提示词效果也不好
image1303×827 317 KB
还是非常明显,推测是风格相关提示词的问题
【High-end anime key visual, Japanese commercial illustration style, ultra-clean lineart with refined cel-shading, soft gradient rendering, glossy highlights, cinematic depth, warm golden lighting palette, polished studio-quality finish, modern anime promotional artwork】,这句比较关键
--【拾壹】--:
我在抖音上也看到有人说会有破碎拼图感,可能是一些新策略导致的副作用吧
--【拾贰】--:
我就是给的参考图,第一张出来一般都没问题,让他改一两次就又出现这种效果了
--【拾叁】--:
那现在就不只宽图了,4:3也会出现问题,而且细节越多的图,出现问题的机率就越大
--【拾肆】--:
满血版只能调用api吗使用吗,我现在直接是在网页端用的
--【拾伍】--:
看起来好像不只是二次元图片会有这个问题,之前看别人发的一些偏现实的也有这个问题,出来的图就好像 Stable Diffusion 采样器和步数设置的有问题一样
--【拾陆】--:
都这样 尤其是前几天到处发的海报,这种破碎纹理感几乎遍布图片,看起来一直觉得怪异
一直也没见什么人提
--【拾柒】--:
体感似乎是提示词和参考图的问题
没有风格指定就默认上碎玻璃风格
有风格指定就把这个覆盖掉了
--【拾捌】--:
很常见,有种马赛克模糊的感觉。我看到别人出的图也类似
--【拾玖】--:
以前的gpt image 也有类似问题,这次特别多。非二次元风格也遇到很多次,感觉是调用了量化版的模型。
有时候不得不去香蕉里洗一下图。

