Implemented inpainting mode

discoart/create.py

@@ -47,6 +47,8 @@ def create(
     init_document: Optional[Union['Document', 'DocumentArray']] = None,
     init_image: Optional[str] = None,
     init_scale: Optional[Union[int, str]] = 1000,
+    inpaint_image: Optional[str] = None,
+    inpaint_mask: Optional[str] = None,
     n_batches: Optional[int] = 4,
     name_docarray: Optional[str] = None,
     on_misspelled_token: Optional[str] = 'ignore',
@@ -123,6 +125,8 @@ def create(**kwargs) -> Optional['DocumentArray']:
     :param init_document: [DiscoArt] Use a Document object as the initial state for DD: its ``.tags`` will be used as parameters, ``.uri`` (if present) will be used as init image.
     :param init_image: Recall that in the image sequence above, the first image shown is just noise.  If an init_image is provided, diffusion will replace the noise with the init_image as its starting state.  To use an init_image, upload the image to the Colab instance or your Google Drive, and enter the full image path here. If using an init_image, you may need to increase skip_steps to ~ 50% of total steps to retain the character of the init. See skip_steps above for further discussion.
     :param init_scale: This controls how strongly CLIP will try to match the init_image provided.  This is balanced against the clip_guidance_scale (CGS) above.  Too much init scale, and the image won’t change much during diffusion. Too much CGS and the init image will be lost.[DiscoArt] Can be scheduled via syntax `[val1]*400+[val2]*600`.
+    :param inpaint_image: [DiscoArt] If an inpaint image and mask is provided, final image will match the inpaint image in non-masked areas.
+    :param inpaint_mask: [DiscoArt] If an inpaint image and mask is provided, final image will match the inpaint image in non-masked areas.
     :param n_batches: This variable sets the number of still images you want DD to create.  If you are using an animation mode (see below for details) DD will ignore n_batches and create a single set of animated frames based on the animation settings.
     :param name_docarray: [DiscoArt] When specified, it overrides the default naming schema of the resulted DocumentArray. Useful when you have to know the result DocumentArray name in advance.The name also supports variable substitution via `{}`. For example, `name_docarray='test-{steps}-{perlin_init}'` will give the name of the DocumentArray as `test-250-False`. Any variable in the config can be substituted.
     :param on_misspelled_token: [DiscoArt] Strategy when encounter misspelled token, can be 'raise', 'correct' and 'ignore'. If 'raise', then the misspelled token in the prompt will raise a ValueError. If 'correct', then the token will be replaced with the correct token. If 'ignore', then the token will be ignored but a warning will show.

discoart/nn/inpainting.py

@@ -0,0 +1,13 @@
+def wrap_inpaint(diff, inpaint_tensor, mask_tensor, noise):
+    def inpaint_sampling(f):
+        def func(*args, **kwargs):
+            result = f(*args, **kwargs)
+            curr_t = args[-1]
+            curr_image = diff.q_sample(inpaint_tensor, curr_t, noise)
+            result["sample"] = curr_image * mask_tensor + result["sample"] * (1 - mask_tensor)
+            return result
+
+        return func
+
+    diff.ddim_sample = inpaint_sampling(diff.ddim_sample)
+    diff.plms_sample = inpaint_sampling(diff.plms_sample)

discoart/resources/default.yml

@@ -4,6 +4,8 @@ text_prompts:
 
 init_image:
 width_height: [ 1280, 768 ]
+inpaint_image:
+inpaint_mask:
 
 skip_steps: 0
 steps: 250

discoart/runner.py

@@ -32,6 +32,7 @@
 from .nn.transform import symmetry_transformation_fn
 from .persist import _sample_thread, _persist_thread, _save_progress_thread
 from .prompt import PromptPlanner
+from .nn.inpainting import wrap_inpaint
 
 
 def do_run(args, models, device, events) -> 'DocumentArray':
@@ -355,6 +356,14 @@ def cond_fn(x, t, **kwargs):
 
         return r_grad
 
+    if args.inpaint_image and args.inpaint_mask:
+        d = Document(uri=args.inpaint_image).load_uri_to_image_tensor(side_x, side_y)
+        inpaint = TF.to_tensor(d.tensor).to(device).unsqueeze(0).mul(2).sub(1)
+        d = Document(uri=args.inpaint_mask).load_uri_to_image_tensor(side_x, side_y)
+        mask = TF.to_tensor(d.tensor).to(device).unsqueeze(0)
+        noise = torch.randn(*inpaint.size(), device=device)
+        wrap_inpaint(diffusion, inpaint, mask, noise)
+
     if args.diffusion_sampling_mode == 'ddim':
         sample_fn = diffusion.ddim_sample_loop_progressive
     else: