Merge branch 'master' into assets-redo

Author: Kosinkadink

.github/workflows/test-ci.yml

@@ -5,6 +5,7 @@ on:
   push:
     branches:
       - master
+      - release/**
     paths-ignore:
       - 'app/**'
       - 'input/**'

.github/workflows/test-execution.yml

@@ -2,9 +2,9 @@ name: Execution Tests
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:

.github/workflows/test-launch.yml

@@ -2,9 +2,9 @@ name: Test server launches without errors
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:

.github/workflows/test-unit.yml

@@ -2,9 +2,9 @@ name: Unit Tests
 
 on:
   push:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
   pull_request:
-    branches: [ main, master ]
+    branches: [ main, master, release/** ]
 
 jobs:
   test:

.github/workflows/update-version.yml

@@ -6,6 +6,7 @@ on:
       - "pyproject.toml"
     branches:
       - master
+      - release/**
 
 jobs:
   update-version:

comfy/cli_args.py

@@ -97,6 +97,13 @@ class LatentPreviewMethod(enum.Enum):
     Latent2RGB = "latent2rgb"
     TAESD = "taesd"
 
+    @classmethod
+    def from_string(cls, value: str):
+        for member in cls:
+            if member.value == value:
+                return member
+        return None
+
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
 parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")

comfy/context_windows.py

@@ -87,6 +87,7 @@ class IndexListCallbacks:
     COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
     EXECUTE_START = "execute_start"
     EXECUTE_CLEANUP = "execute_cleanup"
+    RESIZE_COND_ITEM = "resize_cond_item"
 
     def init_callbacks(self):
         return {}
@@ -166,6 +167,18 @@ def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: Inde
                         new_cond_item = cond_item.copy()
                         # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
                         for cond_key, cond_value in new_cond_item.items():
+                            # Allow callbacks to handle custom conditioning items
+                            handled = False
+                            for callback in comfy.patcher_extension.get_all_callbacks(
+                                IndexListCallbacks.RESIZE_COND_ITEM, self.callbacks
+                            ):
+                                result = callback(cond_key, cond_value, window, x_in, device, new_cond_item)
+                                if result is not None:
+                                    new_cond_item[cond_key] = result
+                                    handled = True
+                                    break
+                            if handled:
+                                continue
                             if isinstance(cond_value, torch.Tensor):
                                 if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
                                    (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):

comfy/k_diffusion/sampling.py

@@ -1618,6 +1618,17 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
             x = x + sde_noise * sigmas[i + 1] * s_noise
     return x
 
+@torch.no_grad()
+def sample_exp_heun_2_x0(model, x, sigmas, extra_args=None, callback=None, disable=None, solver_type="phi_2"):
+    """Deterministic exponential Heun second order method in data prediction (x0) and logSNR time."""
+    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None, r=1.0, solver_type=solver_type)
+
+
+@torch.no_grad()
+def sample_exp_heun_2_x0_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type="phi_2"):
+    """Stochastic exponential Heun second order method in data prediction (x0) and logSNR time."""
+    return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=1.0, solver_type=solver_type)
+
 
 @torch.no_grad()
 def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
@@ -1765,7 +1776,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
         # Predictor
         if sigmas[i + 1] == 0:
             # Denoising step
-            x = denoised
+            x_pred = denoised
         else:
             tau_t = tau_func(sigmas[i + 1])
             curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
@@ -1786,7 +1797,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
             if tau_t > 0 and s_noise > 0:
                 noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise
                 x_pred = x_pred + noise
-    return x
+    return x_pred
 
 
 @torch.no_grad()

comfy/ldm/lumina/model.py

@@ -634,8 +634,11 @@ def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, trans
         img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options)
         freqs_cis = freqs_cis.to(img.device)
 
+        transformer_options["total_blocks"] = len(self.layers)
+        transformer_options["block_type"] = "double"
         img_input = img
         for i, layer in enumerate(self.layers):
+            transformer_options["block_index"] = i
             img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
             if "double_block" in patches:
                 for p in patches["double_block"]:

comfy/ldm/qwen_image/model.py

@@ -218,9 +218,24 @@ def __init__(
             operations=operations,
         )
 
-    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    def _apply_gate(self, x, y, gate, timestep_zero_index=None):
+        if timestep_zero_index is not None:
+            return y + torch.cat((x[:, :timestep_zero_index] * gate[0], x[:, timestep_zero_index:] * gate[1]), dim=1)
+        else:
+            return torch.addcmul(y, gate, x)
+
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor, timestep_zero_index=None) -> Tuple[torch.Tensor, torch.Tensor]:
         shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
-        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
+        if timestep_zero_index is not None:
+            actual_batch = shift.size(0) // 2
+            shift, shift_0 = shift[:actual_batch], shift[actual_batch:]
+            scale, scale_0 = scale[:actual_batch], scale[actual_batch:]
+            gate, gate_0 = gate[:actual_batch], gate[actual_batch:]
+            reg = torch.addcmul(shift.unsqueeze(1), x[:, :timestep_zero_index], 1 + scale.unsqueeze(1))
+            zero = torch.addcmul(shift_0.unsqueeze(1), x[:, timestep_zero_index:], 1 + scale_0.unsqueeze(1))
+            return torch.cat((reg, zero), dim=1), (gate.unsqueeze(1), gate_0.unsqueeze(1))
+        else:
+            return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
 
     def forward(
         self,
@@ -229,14 +244,19 @@ def forward(
         encoder_hidden_states_mask: torch.Tensor,
         temb: torch.Tensor,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        timestep_zero_index=None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         img_mod_params = self.img_mod(temb)
+
+        if timestep_zero_index is not None:
+            temb = temb.chunk(2, dim=0)[0]
+
         txt_mod_params = self.txt_mod(temb)
         img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
         txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
 
-        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1)
+        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1, timestep_zero_index)
         del img_mod1
         txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1)
         del txt_mod1
@@ -251,15 +271,15 @@ def forward(
         del img_modulated
         del txt_modulated
 
-        hidden_states = hidden_states + img_gate1 * img_attn_output
+        hidden_states = self._apply_gate(img_attn_output, hidden_states, img_gate1, timestep_zero_index)
         encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
         del img_attn_output
         del txt_attn_output
         del img_gate1
         del txt_gate1
 
-        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2)
-        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
+        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2, timestep_zero_index)
+        hidden_states = self._apply_gate(self.img_mlp(img_modulated2), hidden_states, img_gate2, timestep_zero_index)
 
         txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2)
         encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
@@ -302,6 +322,7 @@ def __init__(
         pooled_projection_dim: int = 768,
         guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+        default_ref_method="index",
         image_model=None,
         final_layer=True,
         dtype=None,
@@ -314,6 +335,7 @@ def __init__(
         self.in_channels = in_channels
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
+        self.default_ref_method = default_ref_method
 
         self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope))
 
@@ -341,6 +363,9 @@ def __init__(
             for _ in range(num_layers)
         ])
 
+        if self.default_ref_method == "index_timestep_zero":
+            self.register_buffer("__index_timestep_zero__", torch.tensor([]))
+
         if final_layer:
             self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
             self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
@@ -391,11 +416,14 @@ def _forward(
         hidden_states, img_ids, orig_shape = self.process_img(x)
         num_embeds = hidden_states.shape[1]
 
+        timestep_zero_index = None
         if ref_latents is not None:
             h = 0
             w = 0
             index = 0
-            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            ref_method = kwargs.get("ref_latents_method", self.default_ref_method)
+            index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero")
+            timestep_zero = ref_method == "index_timestep_zero"
             for ref in ref_latents:
                 if index_ref_method:
                     index += 1
@@ -415,6 +443,10 @@ def _forward(
                 kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 hidden_states = torch.cat([hidden_states, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+            if timestep_zero:
+                if index > 0:
+                    timestep = torch.cat([timestep, timestep * 0], dim=0)
+                    timestep_zero_index = num_embeds
 
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
@@ -446,7 +478,7 @@ def _forward(
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], timestep_zero_index=timestep_zero_index, transformer_options=args["transformer_options"])
                     return out
                 out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 hidden_states = out["img"]
@@ -458,6 +490,7 @@ def block_wrap(args):
                     encoder_hidden_states_mask=encoder_hidden_states_mask,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    timestep_zero_index=timestep_zero_index,
                     transformer_options=transformer_options,
                 )
 
@@ -474,6 +507,9 @@ def block_wrap(args):
                     if add is not None:
                         hidden_states[:, :add.shape[1]] += add
 
+        if timestep_zero_index is not None:
+            temb = temb.chunk(2, dim=0)[0]
+
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)