SDPose: resize input always (#13349)

.ci/windows_intel_base_files/run_intel_gpu.bat

@@ -0,0 +1,2 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+pause

blueprints/Brightness and Contrast.json

@@ -182,7 +182,7 @@
               ]
             },
             "widgets_values": [
-              50
+              0
             ]
           },
           {

blueprints/Glow.json

@@ -316,7 +316,7 @@
               "step": 1
             },
             "widgets_values": [
-              30
+              0
             ]
           },
           {

comfy/ldm/modules/sdpose.py

@@ -90,7 +90,7 @@ class HeatmapHead(torch.nn.Module):
                 origin_max = np.max(hm[k])
                 dr = np.zeros((H + 2 * border, W + 2 * border), dtype=np.float32)
                 dr[border:-border, border:-border] = hm[k].copy()
-                dr = gaussian_filter(dr, sigma=2.0)
+                dr = gaussian_filter(dr, sigma=2.0, truncate=2.5)
                 hm[k] = dr[border:-border, border:-border].copy()
                 cur_max = np.max(hm[k])
                 if cur_max > 0:

comfy_api/input/__init__.py

@@ -9,7 +9,6 @@ from comfy_api.latest._input import (
     CurveInput,
     MonotoneCubicCurve,
     LinearCurve,
-    RangeInput,
 )
 
 __all__ = [
@@ -22,5 +21,4 @@ __all__ = [
     "CurveInput",
     "MonotoneCubicCurve",
     "LinearCurve",
-    "RangeInput",
 ]

comfy_api/latest/_input/__init__.py

@@ -1,6 +1,5 @@
 from .basic_types import ImageInput, AudioInput, MaskInput, LatentInput
 from .curve_types import CurvePoint, CurveInput, MonotoneCubicCurve, LinearCurve
-from .range_types import RangeInput
 from .video_types import VideoInput
 
 __all__ = [
@@ -13,5 +12,4 @@ __all__ = [
     "CurveInput",
     "MonotoneCubicCurve",
     "LinearCurve",
-    "RangeInput",
 ]

comfy_api/latest/_input/range_types.py

@@ -1,70 +0,0 @@
-from __future__ import annotations
-
-import logging
-import math
-import numpy as np
-
-logger = logging.getLogger(__name__)
-
-
-class RangeInput:
-    """Represents a levels/range adjustment: input range [min, max] with
-    optional midpoint (gamma control).
-
-    Generates a 1D LUT identical to GIMP's levels mapping:
-        1. Normalize input to [0, 1] using [min, max]
-        2. Apply gamma correction: pow(value, 1/gamma)
-        3. Clamp to [0, 1]
-
-    The midpoint field is a position in [0, 1] representing where the
-    midtone falls within [min, max]. It maps to gamma via:
-        gamma = -log2(midpoint)
-    So midpoint=0.5 → gamma=1.0 (linear).
-    """
-
-    def __init__(self, min_val: float, max_val: float, midpoint: float | None = None):
-        self.min_val = min_val
-        self.max_val = max_val
-        self.midpoint = midpoint
-
-    @staticmethod
-    def from_raw(data) -> RangeInput:
-        if isinstance(data, RangeInput):
-            return data
-        if isinstance(data, dict):
-            return RangeInput(
-                min_val=float(data.get("min", 0.0)),
-                max_val=float(data.get("max", 1.0)),
-                midpoint=float(data["midpoint"]) if data.get("midpoint") is not None else None,
-            )
-        raise TypeError(f"Cannot convert {type(data)} to RangeInput")
-
-    def to_lut(self, size: int = 256) -> np.ndarray:
-        """Generate a float64 lookup table mapping [0, 1] input through this
-        levels adjustment.
-
-        The LUT maps normalized input values (0..1) to output values (0..1),
-        matching the GIMP levels formula.
-        """
-        xs = np.linspace(0.0, 1.0, size, dtype=np.float64)
-
-        in_range = self.max_val - self.min_val
-        if abs(in_range) < 1e-10:
-            return np.where(xs >= self.min_val, 1.0, 0.0).astype(np.float64)
-
-        # Normalize: map [min, max] → [0, 1]
-        result = (xs - self.min_val) / in_range
-        result = np.clip(result, 0.0, 1.0)
-
-        # Gamma correction from midpoint
-        if self.midpoint is not None and self.midpoint > 0 and self.midpoint != 0.5:
-            gamma = max(-math.log2(self.midpoint), 0.001)
-            inv_gamma = 1.0 / gamma
-            mask = result > 0
-            result[mask] = np.power(result[mask], inv_gamma)
-
-        return result
-
-    def __repr__(self) -> str:
-        mid = f", midpoint={self.midpoint}" if self.midpoint is not None else ""
-        return f"RangeInput(min={self.min_val}, max={self.max_val}{mid})"

comfy_api/latest/_io.py

@@ -1266,43 +1266,6 @@ class Histogram(ComfyTypeIO):
     Type = list[int]
 
 
-@comfytype(io_type="RANGE")
-class Range(ComfyTypeIO):
-    from comfy_api.input import RangeInput
-    if TYPE_CHECKING:
-        Type = RangeInput
-
-    class Input(WidgetInput):
-        def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None,
-                     socketless: bool=True, default: dict=None,
-                     display: str=None,
-                     gradient_stops: list=None,
-                     show_midpoint: bool=None,
-                     midpoint_scale: str=None,
-                     value_min: float=None,
-                     value_max: float=None,
-                     advanced: bool=None):
-            super().__init__(id, display_name, optional, tooltip, None, default, socketless, None, None, None, None, advanced)
-            if default is None:
-                self.default = {"min": 0.0, "max": 1.0}
-            self.display = display
-            self.gradient_stops = gradient_stops
-            self.show_midpoint = show_midpoint
-            self.midpoint_scale = midpoint_scale
-            self.value_min = value_min
-            self.value_max = value_max
-
-        def as_dict(self):
-            return super().as_dict() | prune_dict({
-                "display": self.display,
-                "gradient_stops": self.gradient_stops,
-                "show_midpoint": self.show_midpoint,
-                "midpoint_scale": self.midpoint_scale,
-                "value_min": self.value_min,
-                "value_max": self.value_max,
-            })
-
-
 DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {}
 def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]):
     DYNAMIC_INPUT_LOOKUP[io_type] = func
@@ -2313,6 +2276,5 @@ __all__ = [
     "BoundingBox",
     "Curve",
     "Histogram",
-    "Range",
     "NodeReplace",
 ]

comfy_api_nodes/nodes_grok.py

@@ -558,7 +558,7 @@ class GrokVideoReferenceNode(IO.ComfyNode):
                 (
                   $res := $lookup(widgets, "model.resolution");
                   $dur := $lookup(widgets, "model.duration");
-                  $refs := inputGroups["model.reference_images"];
+                  $refs := $lookup(inputGroups, "model.reference_images");
                   $rate := $res = "720p" ? 0.07 : 0.05;
                   $price := ($rate * $dur + 0.002 * $refs) * 1.43;
                   {"type":"usd","usd": $price}

comfy_extras/nodes_rtdetr.py

@@ -32,10 +32,12 @@ class RTDETR_detect(io.ComfyNode):
     def execute(cls, model, image, threshold, class_name, max_detections) -> io.NodeOutput:
         B, H, W, C = image.shape
 
-        image_in = comfy.utils.common_upscale(image.movedim(-1, 1), 640, 640, "bilinear", crop="disabled")
-
         comfy.model_management.load_model_gpu(model)
-        results = model.model.diffusion_model(image_in, (W, H))  # list of B dicts
+        results = []
+        for i in range(0, B, 32):
+            batch = image[i:i + 32]
+            image_in = comfy.utils.common_upscale(batch.movedim(-1, 1), 640, 640, "bilinear", crop="disabled")
+            results.extend(model.model.diffusion_model(image_in, (W, H)))
 
         all_bbox_dicts = []
 

comfy_extras/nodes_sdpose.py

@@ -1,5 +1,6 @@
 import torch
 import comfy.utils
+import comfy.model_management
 import numpy as np
 import math
 import colorsys
@@ -410,7 +411,9 @@ class SDPoseDrawKeypoints(io.ComfyNode):
             pose_outputs.append(canvas)
 
         pose_outputs_np = np.stack(pose_outputs) if len(pose_outputs) > 1 else np.expand_dims(pose_outputs[0], 0)
-        final_pose_output = torch.from_numpy(pose_outputs_np).float() / 255.0
+        final_pose_output = torch.from_numpy(pose_outputs_np).to(
+            device=comfy.model_management.intermediate_device(),
+            dtype=comfy.model_management.intermediate_dtype()) / 255.0
         return io.NodeOutput(final_pose_output)
 
 class SDPoseKeypointExtractor(io.ComfyNode):
@@ -459,6 +462,27 @@ class SDPoseKeypointExtractor(io.ComfyNode):
         model_h = int(head.heatmap_size[0]) * 4   # e.g. 192 * 4 = 768
         model_w = int(head.heatmap_size[1]) * 4   # e.g. 256 * 4 = 1024
 
+        def _resize_to_model(imgs):
+            """Aspect-preserving resize + zero-pad BHWC images to (model_h, model_w). Returns (resized_bhwc, scale, pad_top, pad_left)."""
+            h, w = imgs.shape[-3], imgs.shape[-2]
+            scale = min(model_h / h, model_w / w)
+            sh, sw = int(round(h * scale)), int(round(w * scale))
+            pt, pl = (model_h - sh) // 2, (model_w - sw) // 2
+            chw = imgs.permute(0, 3, 1, 2).float()
+            scaled = comfy.utils.common_upscale(chw, sw, sh, upscale_method="bilinear", crop="disabled")
+            padded = torch.zeros(scaled.shape[0], scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
+            padded[:, :, pt:pt + sh, pl:pl + sw] = scaled
+            return padded.permute(0, 2, 3, 1), scale, pt, pl
+
+        def _remap_keypoints(kp, scale, pad_top, pad_left, offset_x=0, offset_y=0):
+            """Remap keypoints from model space back to original image space."""
+            kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
+            invalid = kp[..., 0] < 0
+            kp[..., 0] = (kp[..., 0] - pad_left) / scale + offset_x
+            kp[..., 1] = (kp[..., 1] - pad_top)  / scale + offset_y
+            kp[invalid] = -1
+            return kp
+
         def _run_on_latent(latent_batch):
             """Run one forward pass and return (keypoints_list, scores_list) for the batch."""
             nonlocal captured_feat
@@ -504,36 +528,19 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                         if x2 <= x1 or y2 <= y1:
                             continue
 
-                        crop_h_px, crop_w_px = y2 - y1, x2 - x1
                         crop = img[:, y1:y2, x1:x2, :]  # (1, crop_h, crop_w, C)
-
-                        # scale to fit inside (model_h, model_w) while preserving aspect ratio, then pad to exact model size.
-                        scale = min(model_h / crop_h_px, model_w / crop_w_px)
-                        scaled_h, scaled_w = int(round(crop_h_px * scale)), int(round(crop_w_px * scale))
-                        pad_top, pad_left  = (model_h - scaled_h) // 2, (model_w - scaled_w) // 2
-
-                        crop_chw = crop.permute(0, 3, 1, 2).float()  # BHWC → BCHW
-                        scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="bilinear", crop="disabled")
-                        padded = torch.zeros(1, scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
-                        padded[:, :, pad_top:pad_top + scaled_h, pad_left:pad_left + scaled_w] = scaled
-                        crop_resized = padded.permute(0, 2, 3, 1)  # BCHW → BHWC
+                        crop_resized, scale, pad_top, pad_left = _resize_to_model(crop)
 
                         latent_crop = vae.encode(crop_resized)
                         kp_batch, sc_batch = _run_on_latent(latent_crop)
-                        kp, sc = kp_batch[0], sc_batch[0]  # (K, 2), coords in model pixel space
-
-                        # remove padding offset, undo scale, offset to full-image coordinates.
-                        kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
-                        kp[..., 0] = (kp[..., 0] - pad_left) / scale + x1
-                        kp[..., 1] = (kp[..., 1] - pad_top)  / scale + y1
-
+                        kp = _remap_keypoints(kp_batch[0], scale, pad_top, pad_left, x1, y1)
                         img_keypoints.append(kp)
-                        img_scores.append(sc)
+                        img_scores.append(sc_batch[0])
                 else:
-                    # No bboxes for this image – run on the full image
-                    latent_img = vae.encode(img)
+                    img_resized, scale, pad_top, pad_left = _resize_to_model(img)
+                    latent_img = vae.encode(img_resized)
                     kp_batch, sc_batch = _run_on_latent(latent_img)
-                    img_keypoints.append(kp_batch[0])
+                    img_keypoints.append(_remap_keypoints(kp_batch[0], scale, pad_top, pad_left))
                     img_scores.append(sc_batch[0])
 
                 all_keypoints.append(img_keypoints)
@@ -541,19 +548,16 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                 pbar.update(1)
 
         else: # full-image mode, batched
-            tqdm_pbar = tqdm(total=total_images, desc="Extracting keypoints")
-            for batch_start in range(0, total_images, batch_size):
-                batch_end = min(batch_start + batch_size, total_images)
-                latent_batch = vae.encode(image[batch_start:batch_end])
-
+            for batch_start in tqdm(range(0, total_images, batch_size), desc="Extracting keypoints"):
+                batch_resized, scale, pad_top, pad_left = _resize_to_model(image[batch_start:batch_start + batch_size])
+                latent_batch = vae.encode(batch_resized)
                 kp_batch, sc_batch = _run_on_latent(latent_batch)
 
                 for kp, sc in zip(kp_batch, sc_batch):
-                    all_keypoints.append([kp])
+                    all_keypoints.append([_remap_keypoints(kp, scale, pad_top, pad_left)])
                     all_scores.append([sc])
-                    tqdm_pbar.update(1)
 
-                pbar.update(batch_end - batch_start)
+                pbar.update(len(kp_batch))
 
         openpose_frames = _to_openpose_frames(all_keypoints, all_scores, height, width)
         return io.NodeOutput(openpose_frames)

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.42.8
-comfyui-workflow-templates==0.9.44
+comfyui-frontend-package==1.42.10
+comfyui-workflow-templates==0.9.45
 comfyui-embedded-docs==0.4.3
 torch
 torchsde