feat: SUPIR model support (CORE-17) (#13250)

comfy/ldm/modules/diffusionmodules/openaimodel.py

@@ -34,6 +34,16 @@ class TimestepBlock(nn.Module):
 #This is needed because accelerate makes a copy of transformer_options which breaks "transformer_index"
 def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None, time_context=None, num_video_frames=None, image_only_indicator=None):
     for layer in ts:
+        if "patches" in transformer_options and "forward_timestep_embed_patch" in transformer_options["patches"]:
+            found_patched = False
+            for class_type, handler in transformer_options["patches"]["forward_timestep_embed_patch"]:
+                if isinstance(layer, class_type):
+                    x = handler(layer, x, emb, context, transformer_options, output_shape, time_context, num_video_frames, image_only_indicator)
+                    found_patched = True
+                    break
+            if found_patched:
+                continue
+
         if isinstance(layer, VideoResBlock):
             x = layer(x, emb, num_video_frames, image_only_indicator)
         elif isinstance(layer, TimestepBlock):
@@ -49,15 +59,6 @@ def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, out
         elif isinstance(layer, Upsample):
             x = layer(x, output_shape=output_shape)
         else:
-            if "patches" in transformer_options and "forward_timestep_embed_patch" in transformer_options["patches"]:
-                found_patched = False
-                for class_type, handler in transformer_options["patches"]["forward_timestep_embed_patch"]:
-                    if isinstance(layer, class_type):
-                        x = handler(layer, x, emb, context, transformer_options, output_shape, time_context, num_video_frames, image_only_indicator)
-                        found_patched = True
-                        break
-                if found_patched:
-                    continue
             x = layer(x)
     return x
 
@@ -894,6 +895,12 @@ class UNetModel(nn.Module):
             h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
         h = apply_control(h, control, 'middle')
 
+        if "middle_block_after_patch" in transformer_patches:
+            patch = transformer_patches["middle_block_after_patch"]
+            for p in patch:
+                out = p({"h": h, "x": x, "emb": emb, "context": context, "y": y,
+                         "timesteps": timesteps, "transformer_options": transformer_options})
+                h = out["h"]
 
         for id, module in enumerate(self.output_blocks):
             transformer_options["block"] = ("output", id)
@@ -905,8 +912,9 @@ class UNetModel(nn.Module):
                 for p in patch:
                     h, hsp = p(h, hsp, transformer_options)
 
-            h = th.cat([h, hsp], dim=1)
-            del hsp
+            if hsp is not None:
+                h = th.cat([h, hsp], dim=1)
+                del hsp
             if len(hs) > 0:
                 output_shape = hs[-1].shape
             else:

comfy/ldm/supir/supir_modules.py

@@ -0,0 +1,226 @@
+import torch
+import torch.nn as nn
+
+from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
+from comfy.ldm.modules.diffusionmodules.openaimodel import Downsample, TimestepEmbedSequential, ResBlock, SpatialTransformer
+from comfy.ldm.modules.attention import optimized_attention
+
+
+class ZeroSFT(nn.Module):
+    def __init__(self, label_nc, norm_nc, concat_channels=0, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        ks = 3
+        pw = ks // 2
+
+        self.param_free_norm = operations.GroupNorm(32, norm_nc + concat_channels, dtype=dtype, device=device)
+
+        nhidden = 128
+
+        self.mlp_shared = nn.Sequential(
+            operations.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw, dtype=dtype, device=device),
+            nn.SiLU()
+        )
+        self.zero_mul = operations.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw, dtype=dtype, device=device)
+        self.zero_add = operations.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw, dtype=dtype, device=device)
+
+        self.zero_conv = operations.Conv2d(label_nc, norm_nc, 1, 1, 0, dtype=dtype, device=device)
+        self.pre_concat = bool(concat_channels != 0)
+
+    def forward(self, c, h, h_ori=None, control_scale=1):
+        if h_ori is not None and self.pre_concat:
+            h_raw = torch.cat([h_ori, h], dim=1)
+        else:
+            h_raw = h
+
+        h = h + self.zero_conv(c)
+        if h_ori is not None and self.pre_concat:
+            h = torch.cat([h_ori, h], dim=1)
+        actv = self.mlp_shared(c)
+        gamma = self.zero_mul(actv)
+        beta = self.zero_add(actv)
+        h = self.param_free_norm(h)
+        h = torch.addcmul(h + beta, h, gamma)
+        if h_ori is not None and not self.pre_concat:
+            h = torch.cat([h_ori, h], dim=1)
+        return torch.lerp(h_raw, h, control_scale)
+
+
+class _CrossAttnInner(nn.Module):
+    """Inner cross-attention module matching the state_dict layout of the original CrossAttention."""
+    def __init__(self, query_dim, context_dim, heads, dim_head, dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = dim_head * heads
+        self.heads = heads
+        self.to_q = operations.Linear(query_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_k = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_v = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_out = nn.Sequential(
+            operations.Linear(inner_dim, query_dim, dtype=dtype, device=device),
+        )
+
+    def forward(self, x, context):
+        q = self.to_q(x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+        return self.to_out(optimized_attention(q, k, v, self.heads))
+
+
+class ZeroCrossAttn(nn.Module):
+    def __init__(self, context_dim, query_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        heads = query_dim // 64
+        dim_head = 64
+        self.attn = _CrossAttnInner(query_dim, context_dim, heads, dim_head, dtype=dtype, device=device, operations=operations)
+        self.norm1 = operations.GroupNorm(32, query_dim, dtype=dtype, device=device)
+        self.norm2 = operations.GroupNorm(32, context_dim, dtype=dtype, device=device)
+
+    def forward(self, context, x, control_scale=1):
+        b, c, h, w = x.shape
+        x_in = x
+
+        x = self.attn(
+            self.norm1(x).flatten(2).transpose(1, 2),
+            self.norm2(context).flatten(2).transpose(1, 2),
+        ).transpose(1, 2).unflatten(2, (h, w))
+
+        return x_in + x * control_scale
+
+
+class GLVControl(nn.Module):
+    """SUPIR's Guided Latent Vector control encoder. Truncated UNet (input + middle blocks only)."""
+    def __init__(
+        self,
+        in_channels=4,
+        model_channels=320,
+        num_res_blocks=2,
+        attention_resolutions=(4, 2),
+        channel_mult=(1, 2, 4),
+        num_head_channels=64,
+        transformer_depth=(1, 2, 10),
+        context_dim=2048,
+        adm_in_channels=2816,
+        use_linear_in_transformer=True,
+        use_checkpoint=False,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.model_channels = model_channels
+        time_embed_dim = model_channels * 4
+
+        self.time_embed = nn.Sequential(
+            operations.Linear(model_channels, time_embed_dim, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device),
+        )
+
+        self.label_emb = nn.Sequential(
+            nn.Sequential(
+                operations.Linear(adm_in_channels, time_embed_dim, dtype=dtype, device=device),
+                nn.SiLU(),
+                operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device),
+            )
+        )
+
+        self.input_blocks = nn.ModuleList([
+            TimestepEmbedSequential(
+                operations.Conv2d(in_channels, model_channels, 3, padding=1, dtype=dtype, device=device)
+            )
+        ])
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(num_res_blocks):
+                layers = [
+                    ResBlock(ch, time_embed_dim, 0, out_channels=mult * model_channels,
+                             dtype=dtype, device=device, operations=operations)
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    num_heads = ch // num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, num_head_channels,
+                                           depth=transformer_depth[level], context_dim=context_dim,
+                                           use_linear=use_linear_in_transformer,
+                                           use_checkpoint=use_checkpoint,
+                                           dtype=dtype, device=device, operations=operations)
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+            if level != len(channel_mult) - 1:
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        Downsample(ch, True, out_channels=ch, dtype=dtype, device=device, operations=operations)
+                    )
+                )
+                ds *= 2
+
+        num_heads = ch // num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(ch, time_embed_dim, 0, dtype=dtype, device=device, operations=operations),
+            SpatialTransformer(ch, num_heads, num_head_channels,
+                               depth=transformer_depth[-1], context_dim=context_dim,
+                               use_linear=use_linear_in_transformer,
+                               use_checkpoint=use_checkpoint,
+                               dtype=dtype, device=device, operations=operations),
+            ResBlock(ch, time_embed_dim, 0, dtype=dtype, device=device, operations=operations),
+        )
+
+        self.input_hint_block = TimestepEmbedSequential(
+            operations.Conv2d(in_channels, model_channels, 3, padding=1, dtype=dtype, device=device)
+        )
+
+    def forward(self, x, timesteps, xt, context=None, y=None, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+        emb = self.time_embed(t_emb) + self.label_emb(y)
+
+        guided_hint = self.input_hint_block(x, emb, context)
+
+        hs = []
+        h = xt
+        for module in self.input_blocks:
+            if guided_hint is not None:
+                h = module(h, emb, context)
+                h += guided_hint
+                guided_hint = None
+            else:
+                h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        hs.append(h)
+        return hs
+
+
+class SUPIR(nn.Module):
+    """
+    SUPIR model containing GLVControl (control encoder) and project_modules (adapters).
+    State dict keys match the original SUPIR checkpoint layout:
+      control_model.*           -> GLVControl
+      project_modules.*         -> nn.ModuleList of ZeroSFT/ZeroCrossAttn
+    """
+    def __init__(self, device=None, dtype=None, operations=None):
+        super().__init__()
+
+        self.control_model = GLVControl(dtype=dtype, device=device, operations=operations)
+
+        project_channel_scale = 2
+        cond_output_channels = [320] * 4 + [640] * 3 + [1280] * 3
+        project_channels = [int(c * project_channel_scale) for c in [160] * 4 + [320] * 3 + [640] * 3]
+        concat_channels = [320] * 2 + [640] * 3 + [1280] * 4 + [0]
+        cross_attn_insert_idx = [6, 3]
+
+        self.project_modules = nn.ModuleList()
+        for i in range(len(cond_output_channels)):
+            self.project_modules.append(ZeroSFT(
+                project_channels[i], cond_output_channels[i],
+                concat_channels=concat_channels[i],
+                dtype=dtype, device=device, operations=operations,
+            ))
+
+        for i in cross_attn_insert_idx:
+            self.project_modules.insert(i, ZeroCrossAttn(
+                cond_output_channels[i], concat_channels[i],
+                dtype=dtype, device=device, operations=operations,
+            ))

comfy/ldm/supir/supir_patch.py

@@ -0,0 +1,103 @@
+import torch
+from comfy.ldm.modules.diffusionmodules.openaimodel import Upsample
+
+
+class SUPIRPatch:
+    """
+    Holds GLVControl (control encoder) + project_modules (ZeroSFT/ZeroCrossAttn adapters).
+    Runs GLVControl lazily on first patch invocation per step, applies adapters through
+    middle_block_after_patch, output_block_merge_patch, and forward_timestep_embed_patch.
+    """
+    SIGMA_MAX = 14.6146
+
+    def __init__(self, model_patch, project_modules, hint_latent, strength_start, strength_end):
+        self.model_patch = model_patch           # CoreModelPatcher wrapping GLVControl
+        self.project_modules = project_modules   # nn.ModuleList of ZeroSFT/ZeroCrossAttn
+        self.hint_latent = hint_latent           # encoded LQ image latent
+        self.strength_start = strength_start
+        self.strength_end = strength_end
+        self.cached_features = None
+        self.adapter_idx = 0
+        self.control_idx = 0
+        self.current_control_idx = 0
+        self.active = True
+
+    def _ensure_features(self, kwargs):
+        """Run GLVControl on first call per step, cache results."""
+        if self.cached_features is not None:
+            return
+        x = kwargs["x"]
+        b = x.shape[0]
+        hint = self.hint_latent.to(device=x.device, dtype=x.dtype)
+        if hint.shape[0] != b:
+            hint = hint.expand(b, -1, -1, -1) if hint.shape[0] == 1 else hint.repeat((b + hint.shape[0] - 1) // hint.shape[0], 1, 1, 1)[:b]
+        self.cached_features = self.model_patch.model.control_model(
+            hint, kwargs["timesteps"], x,
+            kwargs["context"], kwargs["y"]
+        )
+        self.adapter_idx = len(self.project_modules) - 1
+        self.control_idx = len(self.cached_features) - 1
+
+    def _get_control_scale(self, kwargs):
+        if self.strength_start == self.strength_end:
+            return self.strength_end
+        sigma = kwargs["transformer_options"].get("sigmas")
+        if sigma is None:
+            return self.strength_end
+        s = sigma[0].item() if sigma.dim() > 0 else sigma.item()
+        t = min(s / self.SIGMA_MAX, 1.0)
+        return t * (self.strength_start - self.strength_end) + self.strength_end
+
+    def middle_after(self, kwargs):
+        """middle_block_after_patch: run GLVControl lazily, apply last adapter after middle block."""
+        self.cached_features = None  # reset from previous step
+        self.current_scale = self._get_control_scale(kwargs)
+        self.active = self.current_scale > 0
+        if not self.active:
+            return {"h": kwargs["h"]}
+        self._ensure_features(kwargs)
+        h = kwargs["h"]
+        h = self.project_modules[self.adapter_idx](
+            self.cached_features[self.control_idx], h, control_scale=self.current_scale
+        )
+        self.adapter_idx -= 1
+        self.control_idx -= 1
+        return {"h": h}
+
+    def output_block(self, h, hsp, transformer_options):
+        """output_block_patch: ZeroSFT adapter fusion replaces cat([h, hsp]). Returns (h, None) to skip cat."""
+        if not self.active:
+            return h, hsp
+        self.current_control_idx = self.control_idx
+        h = self.project_modules[self.adapter_idx](
+            self.cached_features[self.control_idx], hsp, h, control_scale=self.current_scale
+        )
+        self.adapter_idx -= 1
+        self.control_idx -= 1
+        return h, None
+
+    def pre_upsample(self, layer, x, emb, context, transformer_options, output_shape, *args, **kw):
+        """forward_timestep_embed_patch for Upsample: extra cross-attn adapter before upsample."""
+        block_type, _ = transformer_options["block"]
+        if block_type == "output" and self.active and self.cached_features is not None:
+            x = self.project_modules[self.adapter_idx](
+                self.cached_features[self.current_control_idx], x, control_scale=self.current_scale
+            )
+            self.adapter_idx -= 1
+        return layer(x, output_shape=output_shape)
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            self.cached_features = None
+            if self.hint_latent is not None:
+                self.hint_latent = self.hint_latent.to(device_or_dtype)
+        return self
+
+    def models(self):
+        return [self.model_patch]
+
+    def register(self, model_patcher):
+        """Register all patches on a cloned model patcher."""
+        model_patcher.set_model_patch(self.middle_after, "middle_block_after_patch")
+        model_patcher.set_model_output_block_patch(self.output_block)
+        model_patcher.set_model_patch((Upsample, self.pre_upsample), "forward_timestep_embed_patch")

comfy/model_patcher.py

@@ -506,6 +506,10 @@ class ModelPatcher:
     def set_model_noise_refiner_patch(self, patch):
         self.set_model_patch(patch, "noise_refiner")
 
+    def set_model_middle_block_after_patch(self, patch):
+        self.set_model_patch(patch, "middle_block_after_patch")
+
+
     def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs):
         rope_options = self.model_options["transformer_options"].get("rope_options", {})
         rope_options["scale_x"] = scale_x