How do you control different character pose in SDXL when using a reference image? [R][D]
Users are seeking methods to control character poses in SDXL when generating pixel art from a reference image. A common approach involves preprocessing the reference image for a cleaner, pixel-art aesthetic. IP-Adapter is utilized for character appearance, while ControlNet pose/rig conditioning manages the target pose. Experimentation includes using multiple reference images (front, rear, left, right) with pose/rig and depth annotations, providing a separate pose reference through the conditioning pipeline.
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年9月12日 21:45 UTC
收录当时偏移：UTC+02026年9月13日 00:01 UTC
- 发布
- 2026年9月12日 21:45
- 收录
- 2026年9月13日 00:01
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
Hi, I’m working on generating ~128×128 pixel art and trying to generate different poses of the same character.
My current approach is roughly:
Start with a reference image and preprocess it into cleaner/more pixel-art-like data (often removing transparency or setting up fixed number of pallets or descaling) Use IP-Adapter for the character/reference appearance. Use ControlNet pose/rig conditioning to control the target pose. I’m also experimenting with multiple references (front, rear, left, right), with pose/rig and depth annotations. For the target pose, I provide a separate pose reference through the conditioning pipeline.
The problem is that the conditioning often seems to conflict. For example, the model may place an arm according to the ControlNet pose but still DUPLICATE the arm shape/position from the reference, sometimes resulting in strange or duplicated limbs.
I’ve tried adjusting ControlNet/IP-Adapter strength and start/end percentages, which helps somewhat, but I still get a lot of inconsistent behavior. Also try to reinject control net strength in different phases though not so far been helpful
What would be a good way to control pose while making sure the character’s appearance from reference images is somewhat preserved??
I’ am a brokie and can’t really have much resource to train sdxl each time just for each character + poses….