PromptGenV2：仅1G低显存福音！更适合F1的CLIP和T5双通道反推！接近Joy性能与速率平衡反推模型

最新推荐文章于 2025-03-28 09:17:59 发布

小敢摘葡萄

最新推荐文章于 2025-03-28 09:17:59 发布

阅读量1.2k

点赞数 22

文章标签：人工智能 AIGC AI绘画 stable diffusion 计算机视觉音视频大数据

本文链接：https://blog.csdn.net/xiaoganbuaiuk/article/details/143845425

版权

PromptGen-V2简介

在之前的文章介绍过一款提示词准确度接近Joy，但性能更平衡的提示词反推模型Florence-2-large-PromptGen 。今日迎来了再次升级Florence-2-large-PromptGen v2.0，这是基于之前的PromptGen 1.5模型的升级版本，引入了更多的新功能和特性：

• 改进的描述质量：对于<GENERATE_TAGS>、<DETAILED_CAPTION>和<MORE_DETAILED_CAPTION>的描述质量提升。

• 新的**<ANALYZE>指令**：帮助模型更好地理解输入图像的构图。
• 显存效率：相比其他模型，这是一个轻量级的描述模型，使用不超过1G的VRAM，快速生成高质量的图像描述。
• 专为Flux模型设计：适用于T5XXL CLIP和CLIP_L，Miaoshou Tagger的新节点“Flux CLIP Text Encode”消除了为创建描述而运行两个独立标记工具的需求。允许在单次生成中轻松填充两个CLIP，这显著提高使用Flux模型时的速度。
• 混合描述风格：<MIXED_CAPTION>结合了更详细的描述和标签，这对于同时使用T5XXL和CLIP_L的FLUX模型非常有用。MiaoshouTagger ComfyUI中添加了一个新的节点来支持此指令。
• <MIXED_CAPTION_PLUS>：结合混合描述的强大功能与分析。
• 模型地址：https://huggingface.co/MiaoshouAI/Florence-2-large-PromptGen-v2.0
• ComfyUI-Miaoshouai-Tagger：https://github.com/miaoshouai/ComfyUI-Miaoshouai-Tagger

PromptGenv2 ComfyUI体验

更新ComfyUI并搜索插件ComfyUI-Miaoshouai-Tagger安装，模型会在首次运行时候自动下载。

• 插件地址：https://github.com/miaoshouai/ComfyUI-Miaoshouai-Tagger
• transformers版本需要不低于4.38.0
• MiaoshouAI/Florence-2-large-PromptGen-v2.0模型：模型会自动下载，如需手动下载模型，则需要下载整个项目全部文件，并放置ComfyUI/LLM。下载地址：https://huggingface.co/MiaoshouAI/Florence-2-large-PromptGen-v2.0/tree/main

Flux文生图工作流

本文涉及ComfyUI工作流和模型均可在LIBLIBAI上下载或在线运行体验：

• FLUX.1哩布在线可运行-黑暗森林工作室：

https://www.liblib.art/modelinfo/488cd9d58cd4421b9e8000373d7da123

• F.1-绮梦流光-水湄凝香：

https://www.liblib.art/modelinfo/134c6dd95aef48e98a22b24e003e026b

• 工作流-Flux文|图生图+LORA+提示反推一键切换工作流：

https://www.liblib.art/modelinfo/782aacd70f604da39e83368c696a02a8

另外LIBLIBAI已支持本地客户端使用可首页下载体验。

PromptGenv2反推工作流

PromptGenv2反推工作流已上传LIBLIB平台：https://www.liblib.art/modelinfo/fcba1596f88e4a29aef9eb0e043030f6?versionUuid=ac49e877274e47009a4602e23459480d

01.牛逼

T5:

A 3d rendering shoot from a bird's eye view about a small wooden house perched on a rocky island surrounded by lush greenery and a waterfall, with the word "nb" spelled out in large, green grass letters in the center of the island. the island is situated in the middle of the image, surrounded by fluffy white clouds, and the sky is a bright, clear blue with a few wispy clouds scattered throughout. the house has a red roof and a small porch, and there are several trees surrounding it, adding to the natural beauty of the scene. the overall atmosphere is peaceful and serene, with a sense of serenity and tranquility.

clip:

solo, outdoors, sky, day, cloud, blue sky, tree, grass, water, no humans, single tree, waterfall

02.沙雕

T5:

A sandcastle in the foreground with a city skyline in the background. the sandcastle is positioned in the middle of the image, with a view of the city skyline through a hole in a stone wall. the cityscape includes tall buildings, skyscrapers, and a tall television tower, all set against a bright blue sky with fluffy white clouds. the water is calm and reflects the buildings, creating a serene atmosphere. the image has a shallow depth of field, allowing the viewer to focus on the intricate details of the sand castle. the overall effect is one of tranquility and serenity, making it a beautiful and picturesque scene.

clip:

shanghai, china, sky, day, cloud, water, architecture, blue sky, building, tower, architecture details, sandcastle, architecture focus, sand castle

03.3D手办

T5:

3D PVCA photo-realistic shoot from a frontal camera angle about a stylized figurine of a young woman with long blonde hair, wearing a revealing outfit, standing confidently on a wooden surface in a modern indoor setting. the image also shows a soft, blurred background of a blurry landscape with rocks and trees. on the middle of the image, a 1girl, who appears to be in her early twenties, is standing with her hands on her hips, looking directly at the viewer with a smile on her face. she has a slender physique, small breasts, and is wearing a green crop top and a light blue, flowing skirt. her hair is styled in long, wavy locks that cascade down her back. she is wearing gold earrings and has a pair of golden sandals on her feet. the lighting is soft and natural, highlighting her smooth skin and the intricate details of her outfit.

clip:

1girl, solo, long hair, breasts, looking at viewer, smile, skirt, navel, closed mouth, jewelry, bare shoulders, medium breasts, standing, twintails, full body, earrings, indoors, midriff, crop top, lips, feet out of frame, sandals, halterneck, hands on own hips, sarong

04.剪影

T5:

Double exposure silhouetteA digital illustration shoot from a portrait camera angle about a portrait of a young woman with a tree reflected in water at sunset. the image also shows a serene and peaceful atmosphere. on the middle of the image, a 20 years old woman with light skin, brown hair tied in a high ponytail, and blue eyes, who appears to be looking directly at the viewer with a neutral expression, is positioned in the center of the frame. she has a slender physique and bare shoulders, with no clothing or accessories. her hair is styled in a messy yet elegant manner, and her hair color is brown hair. her eye color is blue eyes. she is facing the viewer and her expression is neutral. her lips are closed and her nose is slightly upturned. in front of her, a bare tree with no leaves and no branches is reflected in the calm water. the background is a gradient of warm oranges and pinks, creating a peaceful and serene atmosphere.

clip:

1girl, solo, looking at viewer, brown eyes, hair ornament, closed mouth, upper body, outdoors, nude, water, tree, lips, reflection, orange sky, single hair bun

最后，PromptGenv2在反推效果略逊于Joy反推模型，但仅需要1G远比joy 7G小很多硬件要求，并且秒级快速对反推性能，可以作为快速打标、反推的备选方案之列。特别是对于地显存用户，可作为首选方案。

写在最后

感兴趣的小伙伴，赠送全套AIGC学习资料，包含AI绘画、AI人工智能等前沿科技教程和软件工具，具体看这里。
在这里插入图片描述

AIGC技术的未来发展前景广阔，随着人工智能技术的不断发展，AIGC技术也将不断提高。未来，AIGC技术将在游戏和计算领域得到更广泛的应用，使游戏和计算系统具有更高效、更智能、更灵活的特性。同时，AIGC技术也将与人工智能技术紧密结合，在更多的领域得到广泛应用，对程序员来说影响至关重要。未来，AIGC技术将继续得到提高，同时也将与人工智能技术紧密结合，在更多的领域得到广泛应用。

在这里插入图片描述