本工具基于Ovis2-4B和Ovis2-2B 这2个大模型制作而成。
其实早之前我就有分享过反推提示词工具 moondream，但是我昨天遇到了这个大模型，发现功能更加强大，反推出来的提示更加丰富，更多细节描述。这个工具要求显卡比较低，英伟达显卡6g及以上就可以愉快玩耍了
我用同一张图片反推提示词
下面是Ovis2-4B生成的提示词
The image depicts a young woman standing in a snowy forest. She is wearing a black puffer jacket over a black bikini, creating a striking contrast against the white snow. The woman has long, wavy blonde hair that falls past her shoulders. Her facial expression is neutral, with a slight hint of a smile, and her eyes are looking directly at the camera.
Her posture is relaxed, with her hands casually placed in the pockets of her jacket. The background features snow-covered trees, creating a serene and wintry atmosphere. The lighting appears soft and diffused, typical of an overcast winter day, which enhances the cold and tranquil mood of the scene.
The woman's body shape is slender, and her skin tone is fair. The overall composition of the image, with the contrast between her dark clothing and the white snow, draws attention to her figure and the juxtaposition of her attire against the winter setting. The image captures a moment of contrast between the warmth of her clothing and the cold environment, creating a visually striking and thought-provoking scene.
下面是Ovis2-2B生成的提示词
The image depicts a woman standing in a snowy forest, creating a striking contrast between her attire and the winter setting. She is wearing a black puffer jacket with the hood up, which is open, revealing her black bikini top and black underwear with crisscross straps. Her long, blonde hair cascades down her shoulders, and she has a neutral expression on her face.
The background is a serene winter scene with snow-covered trees, adding to the cold, wintry atmosphere. The lighting is soft and diffused, likely due to overcast skies, which enhances the overall mood of the image.
The woman's posture is relaxed yet confident, with her body slightly turned to the side. Her skin tone is fair, and her body shape appears to be slender. The combination of her black clothing and the snowy environment creates a visually striking contrast, emphasizing her figure and the stark beauty of the winter landscape.
moondream 反推生成的提示词
The image depicts a young woman standing in a snowy forest. She is wearing a black puffy coat and black bikini bottoms. Her blonde hair is styled in loose waves, and she has a neutral expression on her face. The background is filled with snow-covered trees, creating a serene and tranquil atmosphere. The lighting is natural, with sunlight filtering through the trees, casting a soft glow on the scene. The woman's posture is relaxed, with her body angled slightly to the left, and her gaze is directed straight at the camera. The contrast between her black clothing and the white snow adds a dramatic effect to the image. The woman's body shape is slender, and her chest circumference is visible, highlighting her figure.
差距还是有的。所以抽空做了这个反推工具
Ovis2-4B和Ovis2-2B 的区别在于 4b反推的提示词细节更多，但是速度相对慢一点。2b速度快一点，细节没有那么多。我个人觉得其实2b就够用了，但是我觉得4b可能心理感觉更好也就保留下了
因为图片的数量很大的话，那么处理的时间就会差距很大了
简单说下怎么使用。把需要反推的图片放在input 文件夹里面。然后双击启动批量反推图片.exe。选模型1或者2
等待运行完毕就可以了。生成好的提示词文件放在txt文件夹里面。
有了这些大量的提示词后就可以用魔搭 批量AI文生图工具 27pic-api v3.0 无需显卡，无需搭建解压即可运行。来挂机生成大量的图片了。
下面是视频反推出来的提示
1 The video features a woman dressed in a traditional Chinese qipao, a dress with a high collar and a floral pattern, paired with a light yellow blazer. The woman is seen in various poses, with her hair tied back in a ponytail, and she wears long, dangling earrings. The background is a soft pink color with a framed picture visible. Throughout the video, the woman's expressions change subtly, with her eyes looking off to the side and her mouth slightly open, suggesting a range of emotions.
2 The video features a woman in a light blue bikini standing in a hot spring, holding a glass jar. She is surrounded by a bamboo fence and stone edges, creating a serene and natural setting. The woman is seen adjusting her bikini and holding the jar, with water splashing around her. She then holds the jar closer to the camera, showcasing the water inside. The video captures her in various poses, emphasizing the tranquil atmosphere of the hot spring.
3 The video features a woman standing in front of a wooden door, wearing a blue, form-fitting dress with a unique design, one shoulder strap, and a tied belt at the waist. She has long, dark hair and is accessorized with long, matching gloves on one arm. The background is a neutral-colored wall, and the lighting is warm, creating a cozy atmosphere. The woman strikes various poses, showcasing her dress and accessories. Text appears in the top left corner, indicating the source of the video, while the bottom right corner displays the Douyin (TikTok) username '抖音号: 4013595' and the search term '乔乔不熬夜' (Jiao Jiao doesn't stay up late). The video concludes with a screen showing the Douyin profile of the user with the username '抖音号: 4013595' and a search bar with the text '抖音搜索 乔乔不熬夜' (Search on Douyin: Jiao Jiao doesn't stay up late).
4 The video features a woman performing a series of acrobatic movements on a sports field, dressed in a white and light blue crop top and light gray pants. The background includes a building with orange and white walls, pink banners with white Chinese characters, and a few people sitting and walking around. The woman starts by standing and then begins to move, performing a cartwheel and a backflip. She continues with a handstand and a backbend, showcasing her flexibility and balance. The scene is set in a sunny outdoor environment with green grass and trees in the background. The video concludes with the woman lying on the grass, relaxing and smiling, with a cool emoji appearing above her. The text '王六堡' (Wang Ziqi) and '知乎足' (ZhiQie Foot) appear in the top right corner, indicating the woman's social media handle and the context of the video.
如果不知道去哪里找大量的图片来反推，可以试试下面这个
整合包使用说明必看
https://www.myhelen.cn/helen/267.htm
3.0 更新记录
1 部分人启动出错，增加一个环境重新安装批处理文件，如果启动出错可以执行一次
2 修改图片反推工具保存的txt文件名和图片名一致
3 修正部分运行逻辑，速度应该会更快一点
4 清理一部分缓存垃圾的文件 压缩包体积有所减少
5 还是出错就看上面的使用说明
4.0 更新记录
1 更新CUDA 到12.8，理论上已经支持50系显卡了
2 修正部分小BUG
5.0 更新记录 本次更新很大
1 重写处理逻辑。不再区分视频和图片，全部放在input文件夹里面，处理后图片提示词以img开头，视频提示词以video开头。
2 修正以往积累下来的一些bug，处理速度更快。
3 显存8G以下选2B模型，8G以上随意。4B模型推理更强大
视频操作教程
https://www.bilibili.com/video/BV1ScYUzQE3t/?vd_source=f0ca2a91a0d1850ea46d21a82729acaa
File "E:\01-Download\02-Motrix\AI\Ovis-5.0\jian27\lib\json\__init__.py", line 346, in loads
return _default_decoder.decode(s)
File "E:\01-Download\02-Motrix\AI\Ovis-5.0\jian27\lib\json\decoder.py", line 337, in decode
obj, end = self.raw_decode(s, idx=_w(s, 0).end())
File "E:\01-Download\02-Motrix\AI\Ovis-5.0\jian27\lib\json\decoder.py", line 355, in raw_decode
raise JSONDecodeError("Expecting value", s, err.value) from None
json.decoder.JSONDecodeError: Expecting value: line 1
可能存在中文路径或者符号之类
换了两台电脑，一台4070TIS 一台5060Ti 都用不了
提示什么？有仔细看整合包说明？
剑总：我也报错，不支持50系吗？
请问推理的视频建议多少帧，什么参数啊
多少帧 选择哪个模型 取决你的显卡
2B、4B分别的哈
运行报错了，flash_attn实际的版本是2.7.4版本。
AssertionError: Using `flash_attention_2` requires having `flash_attn>=2.6.3` installed.
什么显卡？
Intel(R) UHD Graphics 630,公司工作电脑...
要不 你看看标题再问？
5070 win11 CUDA12.9和12.4都有 程序启动运行正常 运行结束后没有txt文本（无推理结果），图片和视频的反推都可以正常启动运行 但是都没有推理结果 路径也没有中文
看看黑色窗口提示什么
处理进度:100%
1/1 [05:07
请教一下，我跑1.1版本的时候提示进度100%处理完成，然后去看txt文件的时候一个都没有，输入也是英文的文件名（显卡是rtx1660ti，驱动是576，cuda是12.8）
显卡过于上古，不支持
顯卡rtx5060 8g 無法 跑提示詞
看看黑色窗口提示什么
运行 run.py 时出错：Command '['E:\\1111\\jian27\\python.exe', './jian27/run.py']' returned non-zero exit status 1
这是什么问题老大,我用的是WIN10
什么显卡？
文件自带的批处理也运行了,没有任何红字,但是还不行
自行去查 显卡是否支持flash-attn
英伟达3060
可以推理了，仍然无推理结果。
图片和视频都没有TXT文件
用的自带的素材。
晕死
什么显卡？先看看使用说明
File "E:\AI\Ovis2-4B4.0\jian27\lib\site-packages\requests\adapters.py", line 688, in send
raise ConnectTimeout(e, request=request)
requests.exceptions.ConnectTimeout: HTTPSConnectionPool(host='www.modelscope.cn', port=443): Max retries exceeded with url: /api/v1/models/AIDC-AI/Ovis2-2B (Caused by ConnectTimeoutError(, 'Connection to www.modelscope.cn timed out. (connect timeout=60)'))
程序运行失败，错误码：1
如果启动出错就运行一次这个
把你科学上网工具关掉，或者把modelscope.cn加入例外
raise ConnectTimeout(e, request=request)
requests.exceptions.ConnectTimeout: HTTPSConnectionPool(host='www.modelscope.cn', port=443): Max retries exceeded with url: /api/v1/models/AIDC-AI/Ovis2-2B (Caused by ConnectTimeoutError(, 'Connection to www.modelscope.cn timed out. (connect timeout=60)'))
程序运行失败，错误码：1
按Enter键继续...
raise ConnectTimeout(e, request=request)
requests.exceptions.ConnectTimeout: HTTPSConnec
1-4版，我一张图片没推出来
老大，WIN11 无法运行吗？
可以
Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 2/2 [00:00
继续等待
老师再大点的模型怎么下载进去啊?比方说8B 16B 34B？
需要修改代码，没必要下载这么大的模型，效果相差不大