Merge branch 'master' into v3-improvements

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/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/qwen_image/model.py

@@ -322,6 +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,
@@ -334,6 +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))
 
@@ -361,6 +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)
@@ -416,7 +421,7 @@ def _forward(
             h = 0
             w = 0
             index = 0
-            ref_method = kwargs.get("ref_latents_method", "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:

comfy/model_detection.py

@@ -259,7 +259,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_tile_size"] = 512
                 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
-                if "__x0__" in state_dict_keys: # x0 pred
+                if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred
                     dit_config["use_x0"] = True
                 else:
                     dit_config["use_x0"] = False
@@ -618,6 +618,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["image_model"] = "qwen_image"
         dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1]
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
+        if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys:  # 2511
+            dit_config["default_ref_method"] = "index_timestep_zero"
         return dit_config
 
     if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5

comfy/samplers.py

@@ -720,7 +720,7 @@ def max_denoise(self, model_wrap, sigmas):
         sigma = float(sigmas[0])
         return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma
 
-KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
+KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2", "exp_heun_2_x0", "exp_heun_2_x0_sde", "dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",

comfy_api/latest/_io.py

@@ -1594,12 +1594,12 @@ async def EXECUTE_NORMALIZED_ASYNC(cls, *args, **kwargs) -> NodeOutput:
 
     @final
     @classmethod
-    def PREPARE_CLASS_CLONE(cls, v3_data: V3Data) -> type[ComfyNode]:
+    def PREPARE_CLASS_CLONE(cls, v3_data: V3Data | None) -> type[ComfyNode]:
         """Creates clone of real node class to prevent monkey-patching."""
         c_type: type[ComfyNode] = cls if is_class(cls) else type(cls)
         type_clone: type[ComfyNode] = shallow_clone_class(c_type)
         # set hidden
-        type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"])
+        type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"] if v3_data else None)
         return type_clone
 
     @final