RuoyiDu
commented
11 months ago Hi @zzzzzuber, thanks for your interest!
- For the DemoFusion+ControlNet Text2Image sample, I think the main problem is that you need to enter the prompt, otherwise SDXL will generate meaningless noises.
- For the DemoFusion Text2Image sample, there are no obvious problems from the parameters. This is probably a typical failure of DemoFusion, as we recently found out that it's not very good at upscaling images with visible edges (such as comic-style images) and introduces high-frequency noise. Perhaps you could try a simple prompt like "a cute dog" and see if you get good results. If so, you are using it correctly!
zzzzzuber
Thanks for your work, it's very inspration and interesting!!! However, I tried text2img and demofusion + controlnet pipelines, but i cannot obtain very clear and sharp results. Some generate images have grid texture, and they are grid and blury. i don't know if i do something wrong, thanks for your help!!!!!!!! Here are some results about controlnet: image_0,1,2 means different phase
My parameters(same as paper):
prompt = ""
negative_prompt = ""
controlnet_conditioning_scale = 0.5
height=3072
width=3072
num_inference_steps = 50
guidance_scale = 7.5
cosine_scale_1 = 3
cosine_scale_2 = 1
cosine_scale_3 = 1
sigma = 0.8
view_batch_size = 16
stride = 64
seed = 1
Here are some results about text2img:
Parameters:
prompt = "The image features a large building with a prominent advertisement on its side. The advertisement showcases a woman wearing a red dress, and she is holding a bottle of perfume. The building's facade is adorned with a large sign that reads \"BONA.\" The scene captures the attention of passersby with the eye-catching advertisement and the building's distinctive design."
pipe(prompt, negative_prompt=negative_prompt,
height=3072, width=3072, view_batch_size=16, stride=64,
num_inference_steps=50, guidance_scale=7.5,
cosine_scale_1=3, cosine_scale_2=1, cosine_scale_3=1, sigma=0.8,
multi_decoder=True, show_image=False, lowvram=False
)