From 241a1fdb57ffb6f7d1b6dc9e0007dd0c523c808a Mon Sep 17 00:00:00 2001
From: psychedelicious <4822129+psychedelicious@users.noreply.github.com>
Date: Sat, 27 Apr 2024 19:58:46 +1000
Subject: [PATCH] feat(mm): support sdxl ckpt inpainting models

There are only a couple SDXL inpainting models, and my tests indicate they are not as good as SD1.5 inpainting, but at least we support them now.

- Add the config file. This matches what is used in A1111. The only difference from the non-inpainting SDXL config is the number of in-channels.
- Update the legacy config maps to use this config file.
---
 invokeai/backend/model_manager/probe.py       |  1 +
 .../stable-diffusion/sd_xl_inpaint.yaml       | 98 +++++++++++++++++++
 2 files changed, 99 insertions(+)
 create mode 100644 invokeai/configs/stable-diffusion/sd_xl_inpaint.yaml

diff --git a/invokeai/backend/model_manager/probe.py b/invokeai/backend/model_manager/probe.py
index bf21a7fe7b..8f33e4b49f 100644
--- a/invokeai/backend/model_manager/probe.py
+++ b/invokeai/backend/model_manager/probe.py
@@ -51,6 +51,7 @@ LEGACY_CONFIGS: Dict[BaseModelType, Dict[ModelVariantType, Union[str, Dict[Sched
     },
     BaseModelType.StableDiffusionXL: {
         ModelVariantType.Normal: "sd_xl_base.yaml",
+        ModelVariantType.Inpaint: "sd_xl_inpaint.yaml",
     },
     BaseModelType.StableDiffusionXLRefiner: {
         ModelVariantType.Normal: "sd_xl_refiner.yaml",
diff --git a/invokeai/configs/stable-diffusion/sd_xl_inpaint.yaml b/invokeai/configs/stable-diffusion/sd_xl_inpaint.yaml
new file mode 100644
index 0000000000..eea5c15a49
--- /dev/null
+++ b/invokeai/configs/stable-diffusion/sd_xl_inpaint.yaml
@@ -0,0 +1,98 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.13025
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        weighting_config:
+          target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        adm_in_channels: 2816
+        num_classes: sequential
+        use_checkpoint: True
+        in_channels: 9
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [4, 2]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4]
+        num_head_channels: 64
+        use_spatial_transformer: True
+        use_linear_in_transformer: True
+        transformer_depth: [1, 2, 10]  # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16
+        context_dim: 2048
+        spatial_transformer_attn_type: softmax-xformers
+        legacy: False
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          # crossattn cond
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
+            params:
+              layer: hidden
+              layer_idx: 11
+          # crossattn and vector cond
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
+            params:
+              arch: ViT-bigG-14
+              version: laion2b_s39b_b160k
+              freeze: True
+              layer: penultimate
+              always_return_pooled: True
+              legacy: False
+          # vector cond
+          - is_trainable: False
+            input_key: original_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256  # multiplied by two
+          # vector cond
+          - is_trainable: False
+            input_key: crop_coords_top_left
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256  # multiplied by two
+          # vector cond
+          - is_trainable: False
+            input_key: target_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256  # multiplied by two
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
\ No newline at end of file