Unexpected poor test performance

[JCZ404]

Hi, Thanks for the great work! I made a simple evaluation with the provided code, but I encountered some bad performance, and I'm not sure whether there is something wrong with my evaluation. Basically, I curate a new multi-concept combination example with the 'cat' from the FreeCustom/dataset/freecustom/multi_concept/cat_necklace and the 'dog' from the FreeCustom/dataset/freecustom/multi_concept/dog_hat_sunglasses_0. With the provided image and mask, I define a new config as follows:

mark: ""

model_path: "runwayml/stable-diffusion-v1-5"

gpu: 1

# feature injection setting
start_step: 0
end_step: 50
layer_idx: [10,11,12,13,14,15]

# set seeds
seeds: [2]

# set reference image infos
ref_image_infos: {
  dataset/freecustom/multi_concept/cat_dog/image/cat.jpg: a cat,
  dataset/freecustom/multi_concept/cat_dog/image/dog.jpg: a dog,
}
target_prompt: "a cat with a dog"

# ref image setting
use_null_ref_prompts: False     # Whether or not to use an empty prompt when denoising reference images.
mask_weights: [3.0, 3.0]        # The larger the mask_weight is, the higher attention is for the reference image.

negative_prompt: "lowres, bad anatomy, text, error, cropped, worst quality, low quality, normal quality, jpeg artifacts, blurry"
style_fidelity: 1 # The larger the style_fidelity, the more like the reference concepts, range of values: [0,1]

But the final generation result is pretty bad and looks like:

Therefore, I wonder if there is any potential issue for my evaluation or any suggestions to improve the performance?

[wude1234]

权重改一下