模型来自许多优秀模型的合并和lora微调，最终整合了Heptagram模型，本模型来自于开源社区，向所有开源制作者致敬。
我的设备性能不高，演示视频均为540p，我也无法本地对模型进行训练，经过微调和合并获得的本模型在我的设备上效果很好，所以我选择把它发布。
步数设置：4-8
shift video：6-16
shift audio：3
sampler：res_multistep
scheduler:simple
整合了taomate与hyperflow的加速lora。
This model is the result of merging several excellent models and applying LoRA fine-tuning; it ultimately incorporates the Heptagram model. Originating from the open-source community, this project pays tribute to all open-source creators.
Due to my hardware limitations, All demo videos are in 540p, and I am unable to perform model training locally. However, since this fine-tuned and merged model performs well on my setup, I decided to release it.
Steps: 4–8
Shift video: 6–16
Shift audio: 3
Sampler: res_multistep
Scheduler: simple
Integrates acceleration LoRAs from taomate and hyperflow.
Description
本版本音频处理有些瑕疵，但是在480p低分辨率生成任务上表现良好，具体请下载使用以作判断，在我的设备上这个模型可以解决人脸扭曲模糊的问题，转而用一种带有相机噪点、模糊但保留人脸五官特征的风格，对各种风格如游戏、动漫、3D作品生成良好，且十分遵循提示词。
While the audio processing in this version has some flaws, it performs well in 480p low-resolution generation tasks; please download and test it yourself to judge. On my device, this model resolves issues with facial distortion and blurriness, instead producing a style that incorporates camera noise and softness while preserving facial features. It works well across various styles—such as gaming, anime, and 3D art—and adheres closely to prompts.
FAQ
Comments (23)
Please, share the prompt of the sample video with character clothes selection! It godlike!
it isn't t2v if that is what you were thinking lol.
Use Reference mode and provide a few character images.Then, enter the following prompt and wait for the magic to happen.
subject_definitions:
<Subject 1> is a female character defined by <Picture 1>, <Picture 2>, and <Picture 3>. <Picture 1> and <Picture 3> show the character's face. She has short hair. The full-body, skin-tight combat bodysuit shown in <Picture 2> is <Subject 1>'s outfit. She wears no clothing other than the bodysuit. Do not reference any poses or actions from <Picture 1> or <Picture 2>. The character has no hair accessories or ornaments of any kind.
summary:
Generate an entirely new video featuring <Subject 1> as the female character.
The video should have the visual style of actual AAA game footage, including some game UI elements, buttons, and control prompts. Maintain strict character consistency throughout the video. The camera remains completely stationary.
The scene is a character equipment selection interface in a combat preparation room. <Subject 1> stands in the center of the frame, visible from the waist up. Her hairstyle must match the short hairstyle shown in <Picture 1>.
The character cycles through different outfit styles, including the white bodysuit from <Picture 2>. The outfits are displayed and selected through an in-game equipment selection UI. The available outfits, in order, are:
1. Swimsuit.
2. A sexy, revealing, cropped sailor-style school uniform.
3. A short skirt paired with glossy white stockings.
4. A short-sleeved school uniform with shorts.
5. Finally, a nude outfit.
When the character discovers that she is nude, she suddenly becomes embarrassed, covers her body with her arms, and looks directly at the camera.
Each outfit that the character switches to must have a corresponding item in the equipment selection menu. A mouse cursor moves down the equipment list from top to bottom, clicking each item to switch the character's outfit in sequence.
The bottom-left corner of the screen features an Exit button. The bottom-right corner features a Purchase button. The equipment selection UI remains visible throughout the outfit changes.
retention_analysis:
Preserve <Subject 1>'s original character design, facial features, and overall appearance. Do not reuse any backgrounds or poses from the reference images. Do not introduce unrelated characters, objects, or scenes.
detailed_description:
overall_soundscape:
Short, crisp UI sound effects accompany each button click and outfit change. No background music.
non_diegetic_music:
N/A
just save the video in your PC, he has the metadata , then open the video on the menu of COMFYUI. you will get the full workflow.
试了一下，感觉很好。
91 porn是什么lora
这个诡异的命名其实是这个模型的未加速版本
@eatyet 试用了一下感觉真的还挺好用的，感谢作者分享。想请问一下作者这都大概是用了哪些东西融合呀，有10EROS的成分吗，使用下来好像比EROSbeta5更听我的提示词一点。很好奇。
@Kazuhira 主要是融了dasiwa和eros5，后面还从红潮的短剧模型里面融了一些东西
@eatyet 我的设置是8步，ER SDE+beta57，试下来感觉和eros有一个共同的问题就是在视频第一秒开始画面就会明显的变暗，然后试着设置到multistep+simple这个问题反而更明显了，相比之下ERSDE都会好一点，想请问作者大大知道是什么原因吗。
@Kazuhira 我倒是从来没有遇到这个问题，可能是lora设置的问题？如果你没有加任何lora建议按照我的推荐参数跑一版，当然也可能是提示词的问题，ref模式下参考输入和提示词搭配不当是有可能出现这种问题的，具体可以给我发个email，我如果复现了这个问题我就可以解决，👉[email protected]
@eatyet 应该是工作流的问题，我用了作者大大用的那个dasiwa工作流和参数就感觉好多了，
但是想请问这个流中开了latent upscaler之后，视频的第一帧会变的非常模糊像素化的状况会是什么原因呢？换更高分辨率和步数再用latent upscaler也是一样的问题。视频已经上传至评论区。
这是我目前试用过的画质最好的Minimax-H3融合模型，高动态和小脸表现都很好，连屏幕上显示的小字都保持了可读性，之前试过的很多模型小字都是糊成一团。希望作者能继续开发，让音频也更完美。
What architecture?
Ref2V or FL2V ?
Ref2V and Fl2V
The reason the non-merged Turbo version is important is that there are many variations of Turbo, each yielding different generation results; therefore, it is necessary to distinguish whether an issue stems from the Turbo model itself or the checkpoint.Very good work, sir/madam! I love this model!
Compared to the others, this model is amazing in terms of image quality. It performs especially well when you do the first generation at low resolution and then run everything with a single step during latent upscaling. Audio quality matters to me, so I run the first pass at 13 steps. Great work!
可以分享你的示例工作流么，下载的视频拉到画布没有元数据了
建议作者列一下融合了哪些lora。这样避免和个人使用的lora重复。只是建议，别有压力。模型很不错。
快是快，但是慢动作和油腻了，加速lora的通病
Great model, one of my favorites!
Surprisingly good even at 4-steps! Great job! Look forward to V2!