This will be a series of poses as requested.
Description
FAQ
Comments (13)
Regarding the Ass Up pose, is it Qwen itself that seems to always default to an Asian woman? or is the lora trained that way?
Definitely QWEN's base. I actually eliminated any Asian models from my dataset to try to combat that a bit. You really do need to specify race if you want something specific. QWEN is also a bit weird about black people sometimes. I am trying to figure that out as well.
@ColdNorthMenace QWEN是中国阿里巴巴集团旗下的产品，可能数据集来自中国或者亚洲。
@aatrox168 Thank you. We know. I was just inquiring about the training data for this lora. The fact that the lora doesn't add its own bias is a pretty great thing, in my opinion.
@aatrox168 沒錯，這正是我最初的假設。這很有道理。如果翻譯得不好，請見諒，我用的是Google翻譯。
@Jellai I try! Thanks for using my loras!
@ColdNorthMenace 谢谢你的付出，昨天阿里巴巴集团新出了一个模型，效果非常棒，图片质量很高，6GB以下设备都可以流畅运行，10步就能出图。刚刚发布了其中一个版本，后续会支持lora模型，总之效果很棒。建议你去看一下。链接https://comfyanonymous.github.io/ComfyUI_examples/z_image/
@aatrox168 Z-Image is VERY impressive. I've been having so much fun with it the past few days. It's so fast, and the quality seems impossible for that number of parameters. Honestly, while Qwen has better prompt following, Z-Image has better realism than Qwen, and it is less censored.
@Jellai 非常期待他后续的表现😀
I will dive into zimage when the base model releases. I won't be training on the fp8
I, too, would like to see this LoRA get fully.. ..umm.. ..fleshed out. The problem with poses, especially during the training phase, is that trying to get examples from outside sources (meaning you don't have access to a model), is that even trying to find images that utilize the same pose will still result in variations with elements such as camera angle, adherence to the 'exact' pose, and the specifics of the subject (e.g., body weight, proportions, etc.).
If I may offer a suggestion, it may be a good idea to search for images of yoga poses. While yoga may not be sexual, some of the poses could definitely work for some.. ..anatomical.. ..creations. Plus, there are sources for some great yoga images and the poses will be fairly consistent as long as they follow the established format. Here's a few yoga poses you could look up to see what I'm referring to:
- Uttanasana
- Ardha Uttanasana
- Prasarita Padottanasana (Poses A, B, C, and D)
- Padangusthana
- Upavistha Konasana (Pose B)
- Supta Konasana
- Urdhva Dhanurasana
- Karna Pidasana
- Savasana
Additionally, there's always the "Kama Sutra", training photos for physical activities (wrestling, gymnast, ballet, acrobatics, dance variations, etc.), or anything that requires a very specific body pose or stance. Hope that helps.
OH.. ..and one more useful tool that just got added to the Qwen toolbox is that "Qwen Image Edit 2509" (latest release as of this writing), now supports multiple image inputs for IMG2IMG creations, including the ability to use "images of body frameworks" created during runs in workflows the utilize ControlNet and Openpose. NOTE: Qwen 2509 isn't using ControlNet, so you need to create the "image" by screen capturing the OpenPose node (from a non-Qwen run that is using ControlNet with Openpose), then crop it to include only the Openpose skeletal framework, which you can then import into your workflow as a pose guide.
@Unperson I have been playing with 2509 and working on a comfy workflow to rework certain aspects of my dataset. I only have a 4090 so I can't load the full model locally, but when the time comes, Ill do my work on runpod, batch the entire set and sort thru it. Just going to take time and money, both of which I am short on during slow tattoo season.