Using an image connected to dwpose then in slot 3 or slot 4 of TextEncodeQwenImageEditPlus as image results in only black outputs with V5.3
Using an image connected to dwpose then in slot 3 or slot 4 of TextEncodeQwenImageEditPlus as image results in only black outputs with V5.3. Without dwpose the problem is not there. Are more aware of this bug and how to solve it? Works fine with the default Qwen-Image-Edit and also worked before with older versions of AIO V5. Only have V5.3 currently in stored and in use.
The "black output" seems to be caused by various issues that I haven't experienced myself. I've heard of solutions like disabling sage attention, changing drivers or pytorch versions. I'm not sure what the root cause is.
It happens when you have two different images of the same subject, one in image1 and one in image2. For example if image1 has something blocking the face, image2 has nothing blocking the face, You prompt that the image2 face needs to be used. dwpose is in image 3 or 4 slot. You will get either full black or mostly black background. If you disable dwpose and disable the image3. It will work and reconstruct the subject including the face. This means that V5.3 cannot do all these 3 tasks at once (pose the subject in a new pose in the same generation?)
The "black output" seems to be caused by various issues that I haven't experienced myself. I've heard of solutions like disabling sage attention, changing drivers or pytorch versions. I'm not sure what the root cause is.
Thank you no idea what solved it but right now it is working again with a hidden face in image1, with the face in image2 and it creates the perfect subject as it should be. including dwpose in image3. I think it is several variables causing the black screen and I am not sure what it is yet. Just that it is working right now. I do suspect however that the black screen comes when the model is confused or your asking too much of it.