diff --git a/docker-bake.hcl b/docker-bake.hcl
index 64849a2..79ea180 100644
--- a/docker-bake.hcl
+++ b/docker-bake.hcl
@@ -67,6 +67,7 @@ target "core-cuda" {
     cache-to   = ["type=inline"]
     args = {
         RUNTIME = "cuda"
+        TORCH_INDEX = "cu128"
     }
     depends_on = ["runtime-cuda"]
 }
@@ -90,6 +91,7 @@ target "core-cpu" {
     cache-to   = ["type=inline"]
     args = {
         RUNTIME = "cpu"
+        TORCH_INDEX = "cpu"
     }
     depends_on = ["runtime-cpu"]
 }
diff --git a/services/comfy/complete/extra-requirements.txt b/services/comfy/complete/extra-requirements.txt
index b2f6d20..a995efc 100644
--- a/services/comfy/complete/extra-requirements.txt
+++ b/services/comfy/complete/extra-requirements.txt
@@ -23,6 +23,17 @@ dill>=0.3
 # ComfyUI Inspire Pack dependencies
 webcolors>=24.0
 
+# ComfyUI-Lora-Manager dependencies (node is runtime-installed onto the
+# custom_nodes volume, so its imports must already exist in the venv —
+# a missing module here crash-loops ComfyUI at startup)
+natsort>=8.4
+olefile>=0.47
+toml>=0.10
+GitPython>=3.1
+aiosqlite>=0.20
+beautifulsoup4>=4.12
+platformdirs>=4.0
+
 # RMBG and AILab Segmentation
 timm>=1.0
 groundingdino-py>=0.4.0
diff --git a/services/comfy/core/dockerfile.comfy.core b/services/comfy/core/dockerfile.comfy.core
index 7a12d10..b053174 100644
--- a/services/comfy/core/dockerfile.comfy.core
+++ b/services/comfy/core/dockerfile.comfy.core
@@ -21,11 +21,29 @@ RUN umask 022 && \
 # We clone into a temporary directory and move to /app/ComfyUI to prevent
 # permission issues if we were to mount /app entirely in dev modes
 ARG RUNTIME=cuda
+
+# CUDA wheel index for the torch family (torch / torchaudio / torchvision).
+# Defaults to cu128 for the cuda target; the cpu target overrides this to
+# "cpu" via docker-bake.hcl. Future cuda variants (cu130, etc.) can be
+# added by introducing a new bake target with `args = { TORCH_INDEX = "cu130" }`.
+ARG TORCH_INDEX=cu128
+
 RUN --mount=type=cache,target=/app/.cache/pip \
   umask 022 && \
   git clone https://github.com/comfyanonymous/ComfyUI.git /app/ComfyUI && \
-  pip install --extra-index-url https://download.pytorch.org/whl/cu128 \
-    -r /app/ComfyUI/requirements.txt && \
+  # Install torch family from the wheel index ONLY (not --extra-index-url).
+  # --extra-index-url is additive, so pip picks the highest version across
+  # PyPI + the index. When PyPI ships a newer torch than the cu* index
+  # (which happens regularly), pip picks the PyPI build, leaving torch and
+  # torchaudio compiled against different CUDA versions. Using --index-url
+  # replaces PyPI for this command, so resolution is constrained to the
+  # specified index and the CUDA ABI stays consistent across all three.
+  pip install --index-url https://download.pytorch.org/whl/${TORCH_INDEX} \
+    torch torchaudio torchvision && \
+  # Install the rest of ComfyUI's requirements normally. torch, torchaudio,
+  # and torchvision are already installed; pip will see them satisfied and
+  # skip re-resolution.
+  pip install -r /app/ComfyUI/requirements.txt && \
   pip install -r /app/ComfyUI/manager_requirements.txt
 
 # ==================================================================