Sign In

HunYuan_I2V_NFSW_Workflow

250
4.1k
123
Updated: Feb 12, 2025
stylehunyuan
Type
Workflows
Stats
2,725
Reviews
Published
Feb 8, 2025
Base Model
Hunyuan Video
Hash
AutoV2
9BF7FE1141
Tencent Hunyuan is licensed under the Tencent Hunyuan Community License Agreement, Copyright © 2024 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.
Powered by Tencent Hunyuan

请仔细查看下面的说明(下面有中文)

Please review the instructions below carefully

============================================================

This workflow mainly optimizes images to NFSW Video,Added lora block node, optimized parameter settings, and reduced video artifacts, blur and other issues when using certain hunyuan lora.

此工作流主要针对图片到NFSW视频进行了优化,增加了lora block节点,优化参数设置,减少使用某些hunyuan lora时视频伪影,模糊等问题。

I use it myself, if you have any questions you can leave me a message

Check out the original author of the plugin here:

https://github.com/kijai/ComfyUI-HunyuanVideoWrapper

purgevram needs to be installed manually

purgevram需要自己手动安装

https://github.com/T8mars/comfyui-purgevram

Use the new img2vid lora to replace the original version

https://huggingface.co/leapfusion-image2vid-test/image2vid-960x544/blob/main/img2vid544p.safetensors

You can check out the pinned image to see how my video looks like.

All settings are tested by me on my 4090 graphics card.

I want to tell you what I think is a good setting if your graphics card has 24GB video memory (I haven't tried other video memory sizes):

First, you should consider generating longer videos(I don't think you want to watch a 1-second or 2-second video), because the generated video can be enlarged in high definition later, so you need to crop your pictures to be smaller

Some of the resolutions and num_frames I tested, which do not use shared memory(Using shared memory is very slow):

496x496 181

408x496 217

720x576 101

624x624 101

496x896 101

816x560 101

I do not recommend using IP2V, teacache, hyvid_cfg, etc., as they will consume additional video memory. You should use the video memory based on the resolution and total frame rate of the generated video.

The following are only for NFSW

I don’t recommend using AI to write prompts, try to write them yourself,How to write a prompt word, for example:

anal sex,bouncing breasts,An Asian girl with gigantic breasts is lying on a bed. The man is holding her thighs. The man is moving his hips up and down in fast rhythmic movements, inserting his penis completely into the girl's anus, and then pulling it out in fast pumping movements. The girl's gigantic breasts are bouncing wildly as the man's hips ram into her from behind. The man's penis is very thick and is completely inserted into the girl's asshole. The girl is wearing an unbuttoned shirt and white knee-high socks. The girl rolls her eyes and loses consciousness. The background is a bedroom. High quality.

anal sex,bouncing breasts are Lora's prompts.

The prompt words should include the action and clothing, etc. It is not recommended to write too long, which will distract the weight of the main prompt words,Don’t write any prompts that are irrelevant to your video. Focus on the action description.

In addition, I found that using some loras will affect the quality of the video. It may be that the resolution is low when the loras is trained. I am not sure about the specific reason.

The video resolution should not be lower than 512x512, which will seriously affect the association between the video and the prompt word.

My parameter settings

The parameters I am currently using:

embedded_guidance_scale:12 (If the action is too intense, lower this value)

flow_shirt:12

denoise:1

num_frames:101

steps:30 or 9(Fast version)

LORAs

In addition, you need bouncing breasts lora (mainly to increase the breast shaking effect) and anal sex lora (you can use other ones, I use this one mainly for the effect, but this lora will affect the image quality)

1.

https://civitai.com/models/1052680/dancing-with-breasts-bouncing-hunyuan-video?modelVersionId=1181194

strength:0.85

2.

https://civitai.com/models/1193034/side-anal-sex?modelVersionId=1343233

Strength: 1.1,(The default setting in the workflow is 1.1. If the action is too intense, reduce the weight slightly), the higher the weight, the smoother the action. Note that setting this value requires enabling Lora Block Weight, otherwise there will be serious image quality issues, please check the Lora Block Weight description in the next section.

Lora Block Weight

This node has been added to the new Workflow. Please set it as follows to prevent blurring when using Side Anal Sex Lora. Please see the sample images.

Image selection

The problem of multiple fingers and other defects in the picture is not a big deal.

Don’t choose images with a clear gap between the man and the woman, otherwise the man will not be able to touch the woman’s body when he swings, resulting in insufficient shaking. You should choose images where the man and the woman are close together.

====================================

使用新的img2vid lora 替换原先版本

https://huggingface.co/leapfusion-image2vid-test/image2vid-960x544/blob/main/img2vid544p.safetensors

你可以查看置顶图来查看我的视频效果.

所有设置都是我在我的4090上测试过的。

如果你有24GB显存的话,我觉下下面的设置合适点(其他显存大小我还没试过):

首先,你应该考虑生成更长的视频(我觉得你不会想看1秒或2秒的视频),因为生成的视频后面可以高清放大,所以你需要把你的图片剪裁得更小

我测试的一些分辨率和总帧数,不会使用共享内存(使用共享内存速度会非常慢):

496x496 181

408x496 217

720x576 101

624x624 101

496x896 101

816x560 101

我不建议使用IP2V,teacache,hyvid_cfg等,会消耗额外的显存,你应该把显存用在生成视频的分辨率和总帧数上。

插件的原作者:

https://github.com/kijai/ComfyUI-HunyuanVideoWrapper

以下仅限NFSW

我不建议用AI写提示词,尽量自己写,如何写提示词,举个例子:

anal sex,bouncing breasts,An Asian girl with gigantic breasts is lying on a bed. The man is holding her thighs. The man is moving his hips up and down in fast rhythmic movements, inserting his penis completely into the girl's anus, and then pulling it out in fast pumping movements. The girl's gigantic breasts are bouncing wildly as the man's hips ram into her from behind. The man's penis is very thick and is completely inserted into the girl's asshole. The girl is wearing an unbuttoned shirt and white knee-high socks. The girl rolls her eyes and loses consciousness. The background is a bedroom. High quality.

anal sex,bouncing breasts 是lora的提示词。

提示词要包括动作和服装等,不建议写太长,会分散主提示词的权重,不要写和你视频无关的提示词,动作描述为主。

另外我发现使用部分lora会影响视频质量,可能是训练lora的时候分辨率低,具体原因我也不太清楚。

视频分辨率不建议低于512x512,会严重影响视频和提示词的关联。

我的参数设置

我现在在使用的参数:

embedded_guidance_scale:12 (如果动作太激烈,调低这个数值)

flow_shirt:12

denoise:1

num_frames:101

steps:30 or 9(fast版本)

图片的选择

图片有多手指等瑕疵问题不大.

不要选男女之间有明显空隙的的图片,不然男人摆动的时候会无法接触到女人的身体,导致晃动不够,应该选择男女之间贴住的图.

LORAs

另外需要乳摇lora(主要增加乳房晃动效果)和肛交 lora(你可以用其他的,我用这个主要是效果可以,但这个lora会影响画质).

1.

https://civitai.com/models/1052680/dancing-with-breasts-bouncing-hunyuan-video?modelVersionId=1181194

strength:0.85

2.

https://civitai.com/models/1193034/side-anal-sex?modelVersionId=1343233

strength:1.1,(workflow里默认设置的1.1,如果动作太激烈,稍稍降低权重)权重越高,动作越流畅。注意设置成这个数值需要启用lora block weight,不然会有严重的画质问题,请查看下一节的Lora Block Weight说明。

Lora Block Weight

已经在新的Workflow上增加该节点,请按以下设置防止在使用Side Anal Sex Lora时出现模糊问题,请查看示例图片。