HVision-NKU / StoryDiffusion

Accepted as [NeurIPS 2024] Spotlight Presentation Paper
Apache License 2.0
6k stars 601 forks source link

生成的图片为什么什么出现红色、蓝色的点?batch 内随机采样导致的? #93

Open Amo5 opened 6 months ago

Amo5 commented 6 months ago

image

Amo5 commented 6 months ago

好像一直都有的:放大看,会有红色和蓝色的点 image

Z-YuPeng commented 6 months ago

可以提供一个可复现的参数吗?我这边没有观察到类似现象 image - 2024-05-15T152427 385

Amo5 commented 6 months ago

主要参数是这些: seed = 2024 sd_model_path = models_dict["RealVision"] #"SG161222/RealVisXL_V4.0" general_prompt = "a woman with a white dress" #"a man with a black suit" negative_prompt = "naked, deformed, bad anatomy, disfigured, poorly drawn face, mutation, extra limb, ugly, disgusting, poorly drawn hands, missing limb, floating limbs, disconnected limbs, blurry, watermarks, oversaturated, distorted hands, amputation" prompt_array = ["wake up in the bed", "have breakfast", "is on the road, go to the company", "work in the company", "running in the playground", "reading book in the home" ] style_name = "Photographic" 其他应该都是默认的

Z-YuPeng commented 6 months ago

仍然无法复现,另外你的prompt格式应该是用的旧的code,可以再git pull尝试一下新的代码, python gradio_app_sdxl_specific_id_low_vram.py

Amo5 commented 6 months ago

仍然无法复现,另外你的prompt格式应该是用的旧的code,可以再git pull尝试一下新的代码, python gradio_app_sdxl_specific_id_low_vram.py

我用的是,Comic_Generation.ipynb。 这个文件不是最新的吗?我想本地调试

Z-YuPeng commented 6 months ago

目前 最新的改动 都在 gradio_app_sdxl_specific_id_low_vram.py里面,明天我检查一下ipynb文件

Amo5 commented 6 months ago

目前 最新的改动 都在 gradio_app_sdxl_specific_id_low_vram.py里面,明天我检查一下ipynb文件

好的,希望ipynb也同时更新吧😁

Honlan commented 6 months ago

@Amo5 这个是diffusers sdxl pipeline默认加的watermark

class NoWatermarker:
    def __init__(self):
        pass

    def apply_watermark(self, images: torch.FloatTensor):
        return images

pipe = AutoPipelineForText2Image.from_pretrained(..., add_watermarker=False)
pipe.watermark = NoWatermarker()

这样就没了