Sign In

Water Realflow

433
2.5k
46
Verified:
SafeTensor
Type
LoRA
Stats
2,541
Reviews
Published
May 30, 2023
Base Model
SD 1.5
Trigger Words
realflow
Hash
AutoV2
4EA2D8C1D1

简介 Introduction

这个模型的出发点并不是取代产品图渲染,而是为了还原一些特效广告中的真实水流环绕特效。经过测试,意外发现在二次元模型下调低本模型权重可以获得一些意想不到的效果,例如水中女孩。如果你要支持我,可以通过微信(Anson_NgAu)联系我,让我有电费可以继续跑图训练。

The purpose of this model is not to replace product image rendering, but to recreate realistic water flow effects found in certain special effects advertisements. Surprisingly, through testing, it has been discovered that lowering the weights of this model when combined with a cartoon-style model can yield unexpected results, such as a girl in water. If you would like to support me, you can contact me via WeChat (Anson_NgAu) to contribute towards my electricity expenses for continued training.

请注意,你下载的模型名字和我示例图中的prompts不一样,写prompts时请检查确保名字正确,并记得加上触发词“realflow”!!!

Please note that the name of the model you downloaded may be different from the prompts shown in my example. When writing prompts, please double-check to ensure the correct name and remember to include the trigger word "realflow"!!!

本模型训练时已抹除品牌标签,但需注意的是,产品外形极度相似仍有可能构成侵权嫌疑。如果你用于商用途径,确保使用你自己的产品图作为基础。你可以结合controlnet的reference、图生图或inpainting等功能尽情玩耍!

This model has been trained with brand labels removed. However, it is important to note that if the product design is extremely similar, there is still a possibility of copyright infringement. If you intend to use it for commercial purposes, make sure to use your own product images as a basis. Feel free to explore the functionalities of ControlNet's reference, i2i, or inpainting for added creativity and experimentation!

使用方法:

  • 建议使用chilloutmix生成实物图,权重一般0.6-0.75为宜,不超过0.8(AnythingV3或V5生成水中女孩,权重0.3-0.5)

  • 可以不加花的关键词,或在负面prompts加上“花”,从而只画瓶子和水流

  • 步数建议20-30,采样器使用DPM++ SDE Karras

  • 在512*512,768*768,512*768,768*512可以获得较好效果。不建议开启高清修复!

  • 可以在关键词分别指定背景、瓶子、花的颜色(不常见的颜色可能会造成颜色污染),建议结合xyz图表,改变权重来获得你满意的颜色效果

  • 可以尝试不同关键词来影响瓶子的形状,例如“面霜”、“水瓶”

  • 希望你们能喜欢,并支持我的作品!!!

  • 欢迎评论你使用该模型生成的作品,如果你发现配合哪些其他大模型有更好的效果,请一定不要吝啬,尽管分享!!!

Usage:

  • It is recommended to use "chilloutmix" to generate realistic images, with weights generally ranging from 0.6 to 0.75, but not exceeding 0.8. (For generating a girl in water, you can use "AnythingV3" or "V5" with weights ranging from 0.3 to 0.5.)

  • You can omit specific keyword (flower) or add the word "flower" in negative prompts to focus on drawing only the bottle and water flow.

  • It is suggested to use 20-30 steps with the DPM++ SDE Karras sampler.

  • Good results can be achieved with image sizes of 512x512, 768x768, 512x768, or 768x512. It is not recommended to enable high-resolution refinement.

  • You can specify colors for the background, bottle, and flowers using keywords. However, be aware that using uncommon colors may result in color contamination. It is recommended to consult the XYZ chart and adjust weights to achieve your desired color effects.

  • You can experiment with different keywords to influence the shape of the bottle, such as "face cream" or "water bottle."

  • I hope you enjoy using the model and support my work!!!

  • Feel free to share your generated artwork and provide feedback. If you discover better results by combining this model with other large models, please don't hesitate to share your findings!!!

关于训练:

  • 本模型使用了320张实拍图、处理后的广告图进行训练,尺寸包括512*512,512*768,768*512。并设置了“水”和“瓶子”的reg图集。

  • 这个训练基本上是体力活,90%的时间都是在处理素材,选择好看的瓶子照片很重要。

  • 希望能借此机会抛砖引玉,促成更多此类模型,最终真正帮助可商用的广告设计。

Regarding the training:

  • This model was trained using 320 captured photos and processed advertisement images. The image sizes included 512x512, 512x768, and 768x512. A dedicated REG dataset was created for "water" and "bottle" reference images.

  • Training this model was primarily a labor-intensive task, with approximately 90% of the time spent on processing the materials. Choosing visually appealing bottle photos was crucial.

  • I hope that this endeavor can inspire the development of more models of this kind, ultimately benefiting commercial advertising design in a meaningful way.