Happy 2026 !
The model I'm releasing is mainly for turning people in images into 'invisible person' or 'Living clothes' effects.
This model will be used as a pre-step for training WAN2.2. The Alpha Preview version, now is still based on old dataset, so its capabilities are limited - it's just for initial verification of the effect.
(WAN2.1 concept) https://civarchive.com/models/1650084?modelVersionId=1867692
From current tests, using LoRA gives way better results than the raw QWEN-IMAGE-EDIT-2511 version, but you still need to regenerate/retry 4-6 bunch of times, and it's only good for simple scenes.
On top of that, I accidentally found out that Chinese prompts seem to work even better. No clue why, since all the LoRA training data is in English.
[Alpha Preview version]
You can try using the following positive prompt (Replace the brackets with the target clothing/pants/shoes or whatever else you want to keep), or experiment to find the best approach:
Positive prompt:
图中的人物变成invisible,只看得到漂浮着的{REPLACE HERE},和原版织物形状细节材质完全一致,仿佛被空气般的填充贴合着看不见的身体保持原有的身体曲线.完全看不到身体,4K高清图片,高清晰的织物细节,衣物3D完整性,一致的姿态,一致的光影
Negative prompt:
Bad image, low clarity picture, distorted frame, low quality, inconsistent art style, inconsistent fabric details, inconsistent background, incomplete cuffs, unfinished cuffs, inconsistent pose, logo, text, halo, incorrect image, white outline, incomplete 3D clothing with missing parts, incomplete empty openings, incomplete empty sock openings.
Description
QWEN-IMAGE-EDIT-2511-INVISIBLE v0.1 AlphaPreview
FAQ
Comments (6)
I tested with some tries, the result of use "remove" is better than use "invisible". Lora work Okay on one to two persons in the image. More than that do not work well.
Can you allow onsite generation?
I've never disable onsite generation, but you need to pick the right QWEN-IMAGE-EDIT model, like there's one called 'Qwen_Img_Edit_fp8' that's available.
@qqqq4413 Back in the Day, I couldn't use it. Perhaps Civitai was still checking it, or something. Thanks for the hint and the Lora.
I can't get this to work correctly... I'm believe it's my workflow that's the problem and not the LoRA. I just used an I2I, but it doesn't seem to be working with my current workflow. I may be making a noob mistake...
OK yep, I might be a noob but I'm hella resourceful.. I found a workaround. I just dragged the images below into comfyUi until a workflow showed up. (golden bikini top worked)