Merge branch 'master' into assets-redo

Author: Kosinkadink

.ci/update_windows/update.py

@@ -53,6 +53,16 @@ def pull(repo, remote_name='origin', branch='master'):
     repo.stash(ident)
 except KeyError:
     print("nothing to stash")  # noqa: T201
+except:
+    print("Could not stash, cleaning index and trying again.")  # noqa: T201
+    repo.state_cleanup()
+    repo.index.read_tree(repo.head.peel().tree)
+    repo.index.write()
+    try:
+        repo.stash(ident)
+    except KeyError:
+        print("nothing to stash.")  # noqa: T201
+
 backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S'))
 print("creating backup branch: {}".format(backup_branch_name))  # noqa: T201
 try:

api_server/routes/internal/internal_routes.py

@@ -58,8 +58,13 @@ async def get_files(request: web.Request) -> web.Response:
                 return web.json_response({"error": "Invalid directory type"}, status=400)
 
             directory = get_directory_by_type(directory_type)
+
+            def is_visible_file(entry: os.DirEntry) -> bool:
+                """Filter out hidden files (e.g., .DS_Store on macOS)."""
+                return entry.is_file() and not entry.name.startswith('.')
+
             sorted_files = sorted(
-                (entry for entry in os.scandir(directory) if entry.is_file()),
+                (entry for entry in os.scandir(directory) if is_visible_file(entry)),
                 key=lambda entry: -entry.stat().st_mtime
             )
             return web.json_response([entry.name for entry in sorted_files], status=200)

comfy/k_diffusion/sampling.py

@@ -1557,10 +1557,13 @@ def default_er_sde_noise_scaler(x):
 
 
 @torch.no_grad()
-def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
+def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5, solver_type="phi_1"):
     """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
     arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023)
     """
+    if solver_type not in {"phi_1", "phi_2"}:
+        raise ValueError("solver_type must be 'phi_1' or 'phi_2'")
+
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1600,8 +1603,14 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
         denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
         # Step 2
-        denoised_d = torch.lerp(denoised, denoised_2, fac)
-        x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        if solver_type == "phi_1":
+            denoised_d = torch.lerp(denoised, denoised_2, fac)
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        elif solver_type == "phi_2":
+            b2 = ei_h_phi_2(-h_eta) / r
+            b1 = ei_h_phi_1(-h_eta) - b2
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2)
+
         if inject_noise:
             segment_factor = (r - 1) * h * eta
             sde_noise = sde_noise * segment_factor.exp()

comfy/ldm/hunyuan_video/model.py

@@ -43,6 +43,7 @@ class HunyuanVideoParams:
     meanflow: bool
     use_cond_type_embedding: bool
     vision_in_dim: int
+    meanflow_sum: bool
 
 
 class SelfAttentionRef(nn.Module):
@@ -317,7 +318,7 @@ def forward_orig(
                 timesteps_r = transformer_options['sample_sigmas'][w[0] + 1]
                 timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype)
                 vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype))
-                vec = (vec + vec_r) / 2
+                vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2
 
         if ref_latent is not None:
             ref_latent_ids = self.img_ids(ref_latent)

comfy/ldm/lumina/controlnet.py

@@ -41,6 +41,11 @@ def __init__(
         ffn_dim_multiplier: float = (8.0 / 3.0),
         norm_eps: float = 1e-5,
         qk_norm: bool = True,
+        n_control_layers=6,
+        control_in_dim=16,
+        additional_in_dim=0,
+        broken=False,
+        refiner_control=False,
         dtype=None,
         device=None,
         operations=None,
@@ -49,10 +54,11 @@ def __init__(
         super().__init__()
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
-        self.additional_in_dim = 0
-        self.control_in_dim = 16
+        self.broken = broken
+        self.additional_in_dim = additional_in_dim
+        self.control_in_dim = control_in_dim
         n_refiner_layers = 2
-        self.n_control_layers = 6
+        self.n_control_layers = n_control_layers
         self.control_layers = nn.ModuleList(
             [
                 ZImageControlTransformerBlock(
@@ -74,28 +80,49 @@ def __init__(
         all_x_embedder = {}
         patch_size = 2
         f_patch_size = 1
-        x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * self.control_in_dim, dim, bias=True, device=device, dtype=dtype)
+        x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * (self.control_in_dim + self.additional_in_dim), dim, bias=True, device=device, dtype=dtype)
         all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder
 
+        self.refiner_control = refiner_control
+
         self.control_all_x_embedder = nn.ModuleDict(all_x_embedder)
-        self.control_noise_refiner = nn.ModuleList(
-            [
-                JointTransformerBlock(
-                    layer_id,
-                    dim,
-                    n_heads,
-                    n_kv_heads,
-                    multiple_of,
-                    ffn_dim_multiplier,
-                    norm_eps,
-                    qk_norm,
-                    modulation=True,
-                    z_image_modulation=True,
-                    operation_settings=operation_settings,
-                )
-                for layer_id in range(n_refiner_layers)
-            ]
-        )
+        if self.refiner_control:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    ZImageControlTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        block_id=layer_id,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+        else:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    JointTransformerBlock(
+                        layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        multiple_of,
+                        ffn_dim_multiplier,
+                        norm_eps,
+                        qk_norm,
+                        modulation=True,
+                        z_image_modulation=True,
+                        operation_settings=operation_settings,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
 
     def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input):
         patch_size = 2
@@ -105,9 +132,29 @@ def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input):
         control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
 
         x_attn_mask = None
-        for layer in self.control_noise_refiner:
-            control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input)
+        if not self.refiner_control:
+            for layer in self.control_noise_refiner:
+                control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input)
+
         return control_context
 
+    def forward_noise_refiner_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
+        if self.refiner_control:
+            if self.broken:
+                if layer_id == 0:
+                    return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                if layer_id > 0:
+                    out = None
+                    for i in range(1, len(self.control_layers)):
+                        o, control_context = self.control_layers[i](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+                        if out is None:
+                            out = o
+
+                    return (out, control_context)
+            else:
+                return self.control_noise_refiner[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)
+        else:
+            return (None, control_context)
+
     def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input):
         return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input)

comfy/ldm/lumina/model.py

@@ -536,6 +536,7 @@ def patchify_and_embed(
         bsz = len(x)
         pH = pW = self.patch_size
         device = x[0].device
+        orig_x = x
 
         if self.pad_tokens_multiple is not None:
             pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
@@ -572,13 +573,21 @@ def patchify_and_embed(
 
         freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)
 
+        patches = transformer_options.get("patches", {})
+
         # refine context
         for layer in self.context_refiner:
             cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options)
 
         padded_img_mask = None
-        for layer in self.noise_refiner:
+        x_input = x
+        for i, layer in enumerate(self.noise_refiner):
             x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options)
+            if "noise_refiner" in patches:
+                for p in patches["noise_refiner"]:
+                    out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"})
+                    if "img" in out:
+                        x = out["img"]
 
         padded_full_embed = torch.cat((cap_feats, x), dim=1)
         mask = None
@@ -622,14 +631,15 @@ def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, trans
 
         patches = transformer_options.get("patches", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options)
+        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)
 
+        img_input = img
         for i, layer in enumerate(self.layers):
             img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
             if "double_block" in patches:
                 for p in patches["double_block"]:
-                    out = p({"img": img[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
+                    out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options})
                     if "img" in out:
                         img[:, cap_size[0]:] = out["img"]
                     if "txt" in out:

comfy/model_detection.py

@@ -180,8 +180,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["use_cond_type_embedding"] = False
         if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys:
             dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0]
+            dit_config["meanflow_sum"] = True
         else:
             dit_config["vision_in_dim"] = None
+            dit_config["meanflow_sum"] = False
         return dit_config
 
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
@@ -257,8 +259,10 @@ 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
-                dit_config["use_x0"] = True
+                if "__x0__" in state_dict_keys: # x0 pred
+                    dit_config["use_x0"] = True
+                else:
+                    dit_config["use_x0"] = False
         else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
             dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys

comfy/model_patcher.py

@@ -454,6 +454,9 @@ def set_model_double_block_patch(self, patch):
     def set_model_post_input_patch(self, patch):
         self.set_model_patch(patch, "post_input")
 
+    def set_model_noise_refiner_patch(self, patch):
+        self.set_model_patch(patch, "noise_refiner")
+
     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

comfy/ops.py

@@ -497,15 +497,14 @@ def __init__(
             ) -> None:
                 super().__init__()
 
-                self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype}
-                # self.factory_kwargs = {"device": device, "dtype": dtype}
+                if dtype is None:
+                    dtype = MixedPrecisionOps._compute_dtype
+
+                self.factory_kwargs = {"device": device, "dtype": dtype}
 
                 self.in_features = in_features
                 self.out_features = out_features
-                if bias:
-                    self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs))
-                else:
-                    self.register_parameter("bias", None)
+                self._has_bias = bias
 
                 self.tensor_class = None
                 self._full_precision_mm = MixedPrecisionOps._full_precision_mm
@@ -530,7 +529,14 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                     layer_conf = json.loads(layer_conf.numpy().tobytes())
 
                 if layer_conf is None:
-                    self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False)
+                    dtype = self.factory_kwargs["dtype"]
+                    self.weight = torch.nn.Parameter(weight.to(device=device, dtype=dtype), requires_grad=False)
+                    if dtype != MixedPrecisionOps._compute_dtype:
+                        self.comfy_cast_weights = True
+                    if self._has_bias:
+                        self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=dtype))
+                    else:
+                        self.register_parameter("bias", None)
                 else:
                     self.quant_format = layer_conf.get("format", None)
                     if not self._full_precision_mm:
@@ -560,6 +566,11 @@ def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                         requires_grad=False
                     )
 
+                    if self._has_bias:
+                        self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=MixedPrecisionOps._compute_dtype))
+                    else:
+                        self.register_parameter("bias", None)
+
                     for param_name in qconfig["parameters"]:
                         param_key = f"{prefix}{param_name}"
                         _v = state_dict.pop(param_key, None)
@@ -581,7 +592,7 @@ def state_dict(self, *args, destination=None, prefix="", **kwargs):
                     quant_conf = {"format": self.quant_format}
                     if self._full_precision_mm:
                         quant_conf["full_precision_matrix_mult"] = True
-                    sd["{}comfy_quant".format(prefix)] = torch.frombuffer(json.dumps(quant_conf).encode('utf-8'), dtype=torch.uint8)
+                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
                 return sd
 
             def _forward(self, input, weight, bias):

comfy/quant_ops.py

@@ -399,7 +399,10 @@ def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_roun
         orig_dtype = tensor.dtype
 
         if isinstance(scale, str) and scale == "recalculate":
-            scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max
+            scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max
+            if tensor.dtype not in [torch.float32, torch.bfloat16]:  # Prevent scale from being too small
+                tensor_info = torch.finfo(tensor.dtype)
+                scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max))
 
         if scale is not None:
             if not isinstance(scale, torch.Tensor):