Controlnet inpainting change too much of masking area

[crapthings]

Describe the bug

mask

i got mask with this

def make_inpaint_condition (image, image_mask):
    image = np.array(image.convert("RGB")).astype(np.float32) / 255.0
    image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0

    assert image.shape[0:1] == image_mask.shape[0:1], "image and image_mask must have the same image size"
    image[image_mask > 0.5] = -1.0  # set as masked pixel
    image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
    image = torch.from_numpy(image)
    return image

if use apply mask still bad around edge

output_image = pipe.image_processor.apply_overlay(mask_image1, input_image, output_image)

Because there are too many noise below, the one above cannot cover them.

is it possible to preserve mask area?

---

and below is when i use compel without DiffusersTextualInversionManager

the color changed, but it looks cloth is original

https://github.com/damian0815/compel/issues/86

Reproduction

code from here

https://huggingface.co/docs/diffusers/en/using-diffusers/controlnet

Logs

no log

System Info

0.27.2

Who can help?

No response

[asomoza]

I don't quite understand your whole problem because you're mixing a lot of things in the same issue.

Compel or textual inversion are unrelated to the mask so they don't have anything to do with how the unmasked area is affected.

Also if you're changing the whole background, why are you using a controlnet? which controlnet are you using and why?

I think it all comes to how good or bad your mask is, for example if I use this mask I get this result:

mask	result

Doesn't look too bad, it was a just a quick test without controlnet or anything else, just inpainting with a simple mask. The clothes or the color doesn't change at all because of the mask.

Also just a technicality, but in my opinion when you're changing the whole background is outpainting and not inpainting and to get good results you'll need to do a little more than with inpainting.

[crapthings]

I don't quite understand your whole problem because you're mixing a lot of things in the same issue.

Compel or textual inversion are unrelated to the mask so they don't have anything to do with how the unmasked area is affected.

Also if you're changing the whole background, why are you using a controlnet? which controlnet are you using and why?

I think it all comes to how good or bad your mask is, for example if I use this mask I get this result:

mask result Doesn't look too bad, it was a just a quick test without controlnet or anything else, just inpainting with a simple mask. The clothes or the color doesn't change at all because of the mask.

Also just a technicality, but in my opinion when you're changing the whole background is outpainting and not inpainting and to get good results you'll need to do a little more than with inpainting.

this is text 2 image controlnet pipeline controlnet is inpainting

before i use inpainting model without controlnet inpainting, but still got bad result and finetune inpainting is hard with tool(onetrainer)

how possible to get result like yours

i will try your mask

---

do you use

output_image = pipe.image_processor.apply_overlay(mask_image1, input_image, output_image)

---

i use your mask still

# process mask

def make_inpaint_condition (image, image_mask):
    image = np.array(image.convert("RGB")).astype(np.float32) / 255.0
    image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0

    assert image.shape[0:1] == image_mask.shape[0:1], "image and image_mask must have the same image size"
    image[image_mask > 0.5] = -1.0  # set as masked pixel
    image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
    image = torch.from_numpy(image)
    return image

# pipe

pipe = StableDiffusionControlNetPipeline.from_single_file(
  '/home/zznet/workspace/stable-diffusion-webui/models/Stable-diffusion/realisticVisionV60B1_v51VAE.safetensors',
  torch_dtype = torch.bfloat16,
  controlnet = [
      depth_control,
      inpaint_control,
  ],
)

pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas = True, algorithm_type = 'sde-dpmsolver++', euler_at_final = True)

pipe.enable_freeu(s1 = 0.9, s2 = 0.2, b1 = 1.2, b2 = 1.4)

pipe.to('cuda')

pipe.load_textual_inversion([
  'easynegative.safetensors',
  './FastNegativeV2.pt'
], token = [
  'wtf0',
  'wtf1',
])

# sample

output_image = pipe(
    image = [
        depth_control,
        inpaint_control,
    ],
    prompt = prompt,
    negative_prompt = 'wtf0, wtf1, dof, dummy, mannequin',
    width = renderWidth,
    height = renderHeight,
    num_inference_steps = num_inference_steps,
    guidance_scale = guidance_scale,
    guidance_rescale = 0.5,
    # generator = _generator,
    clip_skip = 0,
    controlnet_conditioning_scale = [
        0.0, # disable for test
        0.999,
    ],
).images[0]

# output_image = pipe.image_processor.apply_overlay(mask_image1, input_image, output_image)

[asomoza]

I'm literally using the the basics:

base = load_image("original.jpeg")
mask = load_image("mask.png")

inpaint = pipe(
    prompt,
    width=1024,
    height=1280,
    image=base,
    mask_image=mask,
    guidance_scale=10,
    strength=0.99,
    num_inference_steps=50,
).images[0]

Also as I don't see the need of controlnet, I'm using just StableDiffusionXLInpaintPipeline. Are you using an inpainting model?

Also I'm using SDXL that's why you see that the quality it's better too, even though I reduced a bit the image size so it doesn't take too long to generate.

Edit: I now see with your edit that you're not using an inpainting model, probably that's your solution.

[crapthings]

re: Edit: I now see with your edit that you're not using an inpainting model, probably that's your solution.

i did use inpainting model, but the mask area still changed, i will try again after i test with sdwebui(text2image + inpainting controlnet), so i switched i did use sd 1.5 inpainting and sdxl inpainting the result is degraded, anyway i will try again

---

sd 1.5 inpainting with controlnet depth map

sd 1.5 inpainting without controlnet depth map

[asomoza]

This is my test switching just the model to SD 1.5

So I think that's the difference between SD 1.5 and SDXL, the model is less capable of blending the background, also this should be expected, I had to lower the resolution to 512x640 which is half of what I used for SDXL.

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.