fix: inpainting now matches photo at high generation strength

- 🎉 fix: inpainted areas correlate with surrounding image, even at 100% generation strength.  Previously if the generation strength was high enough the generated image
would be uncorrelated to the rest of the surrounding image.  It created terrible looking images.
 - fix: mask boundaries are more accurate

README.md

@@ -185,11 +185,14 @@ docker run -it --gpus all -v $HOME/.cache/huggingface:/root/.cache/huggingface -
 ## ChangeLog
 
 **2.0.0**
- - feature: interactive prompt added. access by running `aimg`
- - feature: Specify advanced text based masks using boolean logic and strength modifiers. Mask descriptions must be lowercase. Keywords uppercase.
+ - 🎉 fix: inpainted areas correlate with surrounding image, even at 100% generation strength.  Previously if the generation strength was high enough the generated image
+would be uncorrelated to the rest of the surrounding image.  It created terrible looking images.   
+ - 🎉 feature: interactive prompt added. access by running `aimg`
+ - 🎉 feature: Specify advanced text based masks using boolean logic and strength modifiers. Mask descriptions must be lowercase. Keywords uppercase.
    Valid symbols: `AND`, `OR`, `NOT`, `()`, and mask strength modifier `{+0.1}` where `+` can be any of `+ - * /`. Single character boolean operators also work (`|`, `&`, `!`)
- - feature: apply mask edits to original files with `mask_modify_original` (on by default)
+ - 🎉 feature: apply mask edits to original files with `mask_modify_original` (on by default)
  - feature: auto-rotate images if exif data specifies to do so
+ - fix: mask boundaries are more accurate
  - fix: accept mask images in command line
  - fix: img2img algorithm was wrong and wouldn't at values close to 0 or 1
 
@@ -234,7 +237,6 @@ docker run -it --gpus all -v $HOME/.cache/huggingface:/root/.cache/huggingface -
 
 ## Todo
 
- - refactor how output versions are selected (upscaled, modified original, etc) 
  - performance optimizations
    - ✅ https://github.com/huggingface/diffusers/blob/main/docs/source/optimization/fp16.mdx
    - ✅ https://github.com/CompVis/stable-diffusion/compare/main...Doggettx:stable-diffusion:autocast-improvements#

imaginairy/api.py

@@ -122,13 +122,13 @@ def imagine_image_files(
         prompt = result.prompt
         basefilename = f"{base_count:06}_{prompt.seed}_{prompt.sampler_type}{prompt.steps}_PS{prompt.prompt_strength}_{prompt_normalized(prompt.prompt_text)}"
 
-        for image_type, img in result.images.items():
+        for image_type in result.images:
             subpath = os.path.join(outdir, image_type)
             os.makedirs(subpath, exist_ok=True)
             filepath = os.path.join(
                 subpath, f"{basefilename}_[{image_type}].{output_file_extension}"
             )
-            result.save(filepath)
+            result.save(filepath, image_type=image_type)
             logger.info(f"    🖼  [{image_type}] saved to: {filepath}")
         base_count += 1
         del result
@@ -198,7 +198,12 @@ def imagine(
                     sampler_type = prompt.sampler_type
 
                 sampler = get_sampler(sampler_type, model)
-                mask, mask_image, mask_image_orig = None, None, None
+                mask, mask_image, mask_image_orig, mask_grayscale = (
+                    None,
+                    None,
+                    None,
+                    None,
+                )
                 if prompt.init_image:
                     generation_strength = 1 - prompt.init_image_strength
                     t_enc = int(prompt.steps * generation_strength)
@@ -218,7 +223,7 @@ def imagine(
                     init_image_t = pillow_img_to_torch_image(init_image)
 
                     if prompt.mask_prompt:
-                        mask_image = get_img_mask(
+                        mask_image, mask_grayscale = get_img_mask(
                             init_image, prompt.mask_prompt, threshold=0.1
                         )
                     elif prompt.mask_image:
@@ -239,7 +244,7 @@ def imagine(
                                 mask_image.width // downsampling_factor,
                                 mask_image.height // downsampling_factor,
                             ),
-                            resample=Image.Resampling.NEAREST,
+                            resample=Image.Resampling.LANCZOS,
                         )
                         log_img(mask_image, "latent_mask")
 
@@ -256,9 +261,11 @@ def imagine(
 
                     log_latent(init_latent, "init_latent")
                     # encode (scaled latent)
+                    noise = torch.randn_like(init_latent, device="cpu").to(get_device())
                     z_enc = sampler.stochastic_encode(
                         init_latent,
                         torch.tensor([t_enc - 1]).to(get_device()),
+                        noise=noise,
                     )
                     log_latent(z_enc, "z_enc")
 
@@ -297,12 +304,12 @@ def imagine(
                     x_sample_8_orig = x_sample.astype(np.uint8)
                     img = Image.fromarray(x_sample_8_orig)
                     if mask_image_orig and init_image:
-                        mask_image_orig = mask_image_orig.filter(
+                        mask_final = mask_image_orig.filter(
                             ImageFilter.GaussianBlur(radius=3)
                         )
-                        log_img(mask_image_orig, "reconstituting mask")
-                        mask_image_orig = ImageOps.invert(mask_image_orig)
-                        img = Image.composite(img, init_image, mask_image_orig)
+                        log_img(mask_final, "reconstituting mask")
+                        mask_final = ImageOps.invert(mask_final)
+                        img = Image.composite(img, init_image, mask_final)
                         log_img(img, "reconstituted image")
 
                     upscaled_img = None
@@ -328,7 +335,11 @@ def imagine(
                             upscaled_img = enhance_faces(upscaled_img, fidelity=0.8)
 
                     # put the newly generated patch back into the original, full size image
-                    if prompt.mask_modify_original and mask_image_orig and prompt.init_image:
+                    if (
+                        prompt.mask_modify_original
+                        and mask_image_orig
+                        and prompt.init_image
+                    ):
                         img_to_add_back_to_original = (
                             upscaled_img if upscaled_img else img
                         )
@@ -348,8 +359,8 @@ def imagine(
                         log_img(mask_for_orig_size, "mask for original image size")
 
                         rebuilt_orig_img = Image.composite(
-                            img_to_add_back_to_original,
                             prompt.init_image,
+                            img_to_add_back_to_original,
                             mask_for_orig_size,
                         )
                         log_img(rebuilt_orig_img, "reconstituted original")
@@ -359,7 +370,9 @@ def imagine(
                         prompt=prompt,
                         upscaled_img=upscaled_img,
                         is_nsfw=is_nsfw_img,
-                        modified_original_img=rebuilt_orig_img,
+                        modified_original=rebuilt_orig_img,
+                        mask_binary=mask_image_orig,
+                        mask_grayscale=mask_grayscale,
                     )
 
 

imaginairy/enhancers/clip_masking.py

@@ -1,8 +1,11 @@
 from functools import lru_cache
 from typing import Optional, Sequence
 
+import cv2
+import numpy as np
 import PIL.Image
 import torch
+from kornia.filters import median_blur
 from torchvision import transforms
 
 from imaginairy.img_log import log_img
@@ -41,10 +44,30 @@ def get_img_mask(
     mask_cache = get_img_masks(img, descriptions)
     mask = parsed_mask.apply_masks(mask_cache)
     log_img(mask, "combined mask")
+
+    # try to blur the square shaped artifacts somewhat
+    mask = median_blur(mask.unsqueeze(dim=0).unsqueeze(dim=0), (11, 11)).squeeze()
+    log_img(mask, "median blurred")
+
+    kernel = np.ones((5, 5), np.uint8)
+    mask_g = mask.clone()
+
+    # trial and error shows 0.5 threshold has the best "shape"
     if threshold is not None:
-        mask[mask < threshold] = 0
-        mask[mask >= threshold] = 1
-    return transforms.ToPILImage()(mask)
+        mask[mask < 0.5] = 0
+        mask[mask >= 0.5] = 1
+    log_img(mask, f"mask threshold {0.5}")
+
+    mask_np = mask.cpu().numpy()
+    smoother_strength = 5
+    # grow the mask area to make sure we've masked the thing we care about
+    for _ in range(smoother_strength):
+        mask_np = cv2.dilate(mask_np, kernel)
+    # todo: add an outer blur (not gaussian)
+    mask = torch.from_numpy(mask_np)
+    log_img(mask, "mask after closing (dilation then erosion)")
+
+    return transforms.ToPILImage()(mask), transforms.ToPILImage()(mask_g)
 
 
 def get_img_masks(img, mask_descriptions: Sequence[str]):
@@ -66,7 +89,6 @@ def get_img_masks(img, mask_descriptions: Sequence[str]):
             img.repeat(len(mask_descriptions), 1, 1, 1), mask_descriptions
         )[0]
     preds = transforms.Resize(orig_size)(preds)
-    preds = transforms.GaussianBlur(kernel_size=9)(preds)
 
     preds = [torch.sigmoid(p[0]) for p in preds]
 

imaginairy/samplers/ddim.py

@@ -359,8 +359,12 @@ class DDIMSampler:
                 assert orig_latent is not None
                 xdec_orig = self.model.q_sample(orig_latent, ts)
                 log_latent(xdec_orig, "xdec_orig")
-                log_latent(xdec_orig * mask, "masked_xdec_orig")
-                x_dec = xdec_orig * mask + (1.0 - mask) * x_dec
+                # this helps prevent the weird disjointed images that can happen with masking
+                hint_strength = 0.8
+                xdec_orig_with_hints = (
+                    xdec_orig * (1 - hint_strength) + orig_latent * hint_strength
+                )
+                x_dec = xdec_orig_with_hints * mask + (1.0 - mask) * x_dec
                 log_latent(x_dec, "x_dec")
 
             x_dec, pred_x0 = self.p_sample_ddim(

imaginairy/samplers/plms.py

@@ -388,6 +388,7 @@ class PLMSSampler:
         temperature=1.0,
         mask=None,
         orig_latent=None,
+        noise=None,
     ):
 
         timesteps = self.ddim_timesteps[:t_start]
@@ -398,7 +399,15 @@ class PLMSSampler:
         iterator = tqdm(time_range, desc="PLMS altering image", total=total_steps)
         x_dec = x_latent
         old_eps = []
-
+        log_latent(x_dec, "x_dec")
+
+        # not sure what the downside of using the same noise throughout the process would be...
+        # seems to work fine. maybe it runs faster?
+        noise = (
+            torch.randn_like(x_dec, device="cpu").to(x_dec.device)
+            if noise is None
+            else noise
+        )
         for i, step in enumerate(iterator):
             index = total_steps - i - 1
             ts = torch.full(
@@ -413,10 +422,14 @@ class PLMSSampler:
 
             if mask is not None:
                 assert orig_latent is not None
-                xdec_orig = self.model.q_sample(orig_latent, ts)
+                xdec_orig = self.model.q_sample(orig_latent, ts, noise)
                 log_latent(xdec_orig, "xdec_orig")
-                log_latent(xdec_orig * mask, "masked_xdec_orig")
-                x_dec = xdec_orig * mask + (1.0 - mask) * x_dec
+                # this helps prevent the weird disjointed images that can happen with masking
+                hint_strength = 0.8
+                xdec_orig_with_hints = (
+                    xdec_orig * (1 - hint_strength) + orig_latent * hint_strength
+                )
+                x_dec = xdec_orig_with_hints * mask + (1.0 - mask) * x_dec
                 log_latent(x_dec, "x_dec")
 
             x_dec, pred_x0, e_t = self.p_sample_plms(
@@ -446,5 +459,6 @@ class PLMSSampler:
                 img_callback(pred_x0, "pred_x0")
 
             log_latent(x_dec, f"x_dec {i}")
+            log_latent(x_dec, f"e_t {i}")
             log_latent(pred_x0, f"pred_x0 {i}")
         return x_dec

imaginairy/schema.py

@@ -138,7 +138,6 @@ class ImaginePrompt:
         self.mask_modify_original = mask_modify_original
         self.tile_mode = tile_mode
 
-
     @property
     def prompt_text(self):
         if len(self.prompts) == 1:
@@ -186,7 +185,7 @@ class ImagineResult:
         prompt: ImaginePrompt,
         is_nsfw,
         upscaled_img=None,
-        modified_original_img=None,
+        modified_original=None,
         mask_binary=None,
         mask_grayscale=None,
     ):
@@ -197,8 +196,8 @@ class ImagineResult:
         if upscaled_img:
             self.images["upscaled"] = upscaled_img
 
-        if modified_original_img:
-            self.images["modified_original"] = modified_original_img
+        if modified_original:
+            self.images["modified_original"] = modified_original
 
         if mask_binary:
             self.images["mask_binary"] = mask_binary

tests/test_enhancers.py

@@ -40,8 +40,15 @@ def test_clip_masking():
         "*1",
         "*10",
     ]:
-        pred = get_img_mask(img, f"(head OR face){{{mask_modifier}}}")
-        pred.save(f"{TESTS_FOLDER}/test_output/earring_mask_{mask_modifier}.png")
+        pred_bin, pred_grayscale = get_img_mask(
+            img, f"(head OR face){{{mask_modifier}}}", threshold=0.1
+        )
+        pred_grayscale.save(
+            f"{TESTS_FOLDER}/test_output/earring_mask_{mask_modifier}_g.png"
+        )
+        pred_bin.save(
+            f"{TESTS_FOLDER}/test_output/earring_mask_{mask_modifier}_bin.png"
+        )
 
     prompt = ImaginePrompt(
         "professional photo of a woman",
@@ -57,8 +64,9 @@ def test_clip_masking():
     )
 
     result = next(imagine(prompt))
-    result.modified_original_img.save(
-        f"{TESTS_FOLDER}/test_output/earring_mask_photo.png"
+    result.save(
+        f"{TESTS_FOLDER}/test_output/earring_mask_photo.png",
+        image_type="modified_original",
     )