Dialectical Love - Image-based Video Minimax H3 NSFW Sex Lora:
This is a Lora highly related to SEX. Its training set covers:
Indian white people, white people, yellow people, Arab white people's sexual videos, including: oral sex, basic positions, anal sex, cowboy style, male masturbation, gay male sex.
Athletic or slim, fat or thin, young or old.
Fast or slow, not inserted or inserting, inserting and thrusting.
Male sudden pauses or ejaculation outside the body.
Some animated styles.
X-ray effect (general fitting effect).
Touching breasts.
They do not currently include the following scenarios: female solo masturbation, lesbian sex, transgender. Please help me find the training set. |
I don't yet have the ability to tame the audio. This LORA is the first one I trained for H3. Given the current control of the community over H3 LORAs, this LORA is likely to have many shortcomings. During the generation process, there might be many poor images. The same prompt might require at least 4 generations to determine if it's usable. You need to roll multiple times, and even: the appearance of deformed objects is normal. During the process of constructing the prompt, you can use a short sentence to trigger it. It supports both Chinese and English, for example: "...." "A man is thrusting into the woman's vagina..." /“....... "A woman is performing oral sex on a man..." If these cues don't work in certain sufficiently special situations, you can add some additional descriptions of physical characteristics, such as "The woman turns her back to the camera, the man remains still, and the woman repeatedly squats and stands up vertically with her hips."
It supports tubro accelerated LORA. It performs poorly in REF2V, but you can try adding this LoRa to the workflow of REF2V as a boost with an intensity of 0.1-0.2. Note that if the intensity exceeds 0.2, such as 0.5, the quality of the image will suffer significantly and there may be distorted output.
In the I2V workflow, simply set the intensity to 1.
Please feel free to share your proud works and tick this box. If your work uses this Lora, we would be very grateful.
Overall, this LORA did not meet my expectations. The LORA training for H3 was extremely difficult. If you are willing to share your experience or have a discussion with me, I would be glad to cooperate with you.
Please do not use this LORA to generate any documents that violate the local laws and regulations of your region. These documents are only for learning, communication and sharing purposes, and we assume no responsibility for any possible consequences.
辨证的爱-图片生视频 Minimax H3 nsfw sex lora:
这是一个与SEX高度相关的Lora,它的训练集涵盖了:
印度白种人人 白种人 黄种人 阿拉伯白种人的性爱视频,它们包括:口交丶基础姿势丶肛交丶牛仔丶男性自慰丶男男同性恋。
壮硕或苗条的,胖的或瘦的,年轻或老的。
快速或慢的,未插入或正在插入的,正在抽插的。
男性的突然停顿或体外射精。
部分动画风格。
x-ray效果(拟合效果一般)。
摸乳。
它们暂时未包括如下场合:女性单人自慰丶女女同性恋丶跨性别。欢迎帮助我找到训练集。|
我尚未有能力对音频进行驯化,此lora为我为H3训练的第1个lora,鉴于目前社区对H3lora的掌控力,这个lora很可能也存在许多不足,生成中或许可能会出现很多的差画面,同一个提示词可能需要至少4次生成才能确定是否可用,你需要多次Roll,甚至于:出现畸形的物体都是正常的。
在提示词构建的过程中,你可以使用一段简短的来触发,它支持中文/英文,例如:“.......,男人正在抽插女人的阴道,........”/“.......,女人正在口交男人,........”,如果某些足够特殊的场合没有使这些提示词起效,你可以额外适当地加上一些物理表征的描述,如“女性的背朝向镜头,男性不动,女性的臀部上下垂直反复蹲起”。
它支持 tubro 加速lora。
它在REF2V中表现较差,但你可以尝试在REF2V的工作流中加入此lora,作为0.1-0.2的强度作为助推,注意,如果强度超过0.2,比如0.5,画面的质量将会较大损失,并且可能输出扭曲画面。
在I2V工作流中,直接将强度设置为1即可。
请尽量分享你自豪的作品并勾选本lora,如果它使用了本lora,非常感谢。
总体来说,这个lora总体上没有达到我的预期,因为H3的lora训练非常地困难,如果你愿意分享你的经验或者互相交流,我会乐意配合你。
请勿使用本lora生成违反您当地法律法规的文件,文件仅作学习交流分享,对可能带来的后果不作任何责任。
Description
1.0 6300step
FAQ
Comments (8)
Can it do interracial white/indian couples?
text2video pls
So I love that it works with anal, but the problem is it really degrades the audio. Like an unusable amount for me. Using the standard euler + beta lightx2v lora.
I'm trying out the LORA model for voice merging for the first time. After that, I will attempt to master the training of voices.
@Boccp Its very hard to do so your work is not easy at all.
The current problem facing minimax is the combination of both the video and audio latents processed together at the same time so that really complicates things. everyone appreciates what all the lora makers are attempting so don't give up
@jj43797771 The H3-Omni-Transformer jointly predicts video and audio latents, which are then decoded into video and stereo audio, respectively.
I don't know much about transformer architectures, but that sounds like a terrible idea. What's the advantage of this over separating them and crossconnecting layers, then having the vae decode from the audio only? Better sync? I mean, it beats ltx's irreparably rubbish audio, but still.