diff --git a/installer.py b/installer.py
index ed59080b4..6a8948ff7 100644
--- a/installer.py
+++ b/installer.py
@@ -386,6 +386,9 @@ def pip(arg: str, ignore: bool = False, quiet: bool = True, uv = True):
     t_start = time.time()
     originalArg = arg
     arg = arg.replace('>=', '==')
+    if opts.get('offline_mode', False):
+        log.warning('Offline mode enabled')
+        return
     package = arg.replace("install", "").replace("--upgrade", "").replace("--no-deps", "").replace("--force-reinstall", "").replace(" ", " ").strip()
     uv = uv and args.uv and not package.startswith('git+')
     pipCmd = "uv pip" if uv else "pip"
diff --git a/modules/control/proc/depth_anything/__init__.py b/modules/control/proc/depth_anything/__init__.py
index cf7ee92e2..4709727e4 100644
--- a/modules/control/proc/depth_anything/__init__.py
+++ b/modules/control/proc/depth_anything/__init__.py
@@ -27,7 +27,7 @@ class DepthAnythingDetector:
             PrepareForNet()])
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path: str, cache_dir: str) -> str:
+    def from_pretrained(cls, pretrained_model_or_path: str, cache_dir: str, local_files_only=False) -> str:
         from modules.control.proc.depth_anything.dpt import DPT_DINOv2
         import huggingface_hub as hf
         model = (
@@ -40,7 +40,7 @@ class DepthAnythingDetector:
             .to(devices.device)
             .eval()
         )
-        model_path = hf.hf_hub_download(repo_id=pretrained_model_or_path, filename="pytorch_model.bin", cache_dir=cache_dir)
+        model_path = hf.hf_hub_download(repo_id=pretrained_model_or_path, filename="pytorch_model.bin", cache_dir=cache_dir, local_files_only=local_files_only)
         model_dict = torch.load(model_path)
         model.load_state_dict(model_dict)
         return cls(model)
diff --git a/modules/control/proc/dwpose/__init__.py b/modules/control/proc/dwpose/__init__.py
index e97ade3e6..9e8466a69 100644
--- a/modules/control/proc/dwpose/__init__.py
+++ b/modules/control/proc/dwpose/__init__.py
@@ -51,7 +51,7 @@ def check_dependencies():
     status = [installed(p, reload=False, quiet=True) for p in packages]
     debug(f'DWPose required={packages} status={status}')
     if not all(status):
-        log.info(f'Installing DWPose dependencies: {packages}')
+        log.info(f'Installing dependencies: for=dwpose packages={packages}')
         cmd = 'install --upgrade --no-deps --force-reinstall '
         pkgs = ' '.join(packages)
         pip(cmd + pkgs, ignore=False, quiet=True, uv=False)
diff --git a/modules/control/proc/hed.py b/modules/control/proc/hed.py
index e0144c41f..06610921e 100644
--- a/modules/control/proc/hed.py
+++ b/modules/control/proc/hed.py
@@ -60,12 +60,12 @@ class HEDdetector:
         self.model = model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "ControlNetHED.pth"
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         model = ControlNetHED_Apache2()
         model.load_state_dict(torch.load(model_path, map_location='cpu'))
         model.float().eval()
diff --git a/modules/control/proc/leres/__init__.py b/modules/control/proc/leres/__init__.py
index 63040bc49..e2f358c7d 100644
--- a/modules/control/proc/leres/__init__.py
+++ b/modules/control/proc/leres/__init__.py
@@ -20,13 +20,13 @@ class LeresDetector:
         self.pix2pixmodel = pix2pixmodel
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, pix2pix_filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, pix2pix_filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "res101.pth"
         pix2pix_filename = pix2pix_filename or "latest_net_G.pth"
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         checkpoint = torch.load(model_path, map_location=torch.device('cpu'))
         model = RelDepthModel(backbone='resnext101')
         model.load_state_dict(strip_prefix_if_present(checkpoint['depth_model'], "module."), strict=True)
@@ -34,7 +34,7 @@ class LeresDetector:
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, pix2pix_filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, pix2pix_filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, pix2pix_filename, cache_dir=cache_dir, local_files_only=local_files_only)
         opt = TestOptions().parse()
         if not torch.cuda.is_available():
             opt.gpu_ids = []  # cpu mode
diff --git a/modules/control/proc/lineart.py b/modules/control/proc/lineart.py
index afd55d9a7..61d3ce8b4 100644
--- a/modules/control/proc/lineart.py
+++ b/modules/control/proc/lineart.py
@@ -95,7 +95,7 @@ class LineartDetector:
         self.model_coarse = coarse_model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, coarse_filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, coarse_filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "sk_model.pth"
         coarse_filename = coarse_filename or "sk_model2.pth"
 
@@ -103,8 +103,8 @@ class LineartDetector:
             model_path = os.path.join(pretrained_model_or_path, filename)
             coarse_model_path = os.path.join(pretrained_model_or_path, coarse_filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
-            coarse_model_path = hf_hub_download(pretrained_model_or_path, coarse_filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
+            coarse_model_path = hf_hub_download(pretrained_model_or_path, coarse_filename, cache_dir=cache_dir, local_files_only=local_files_only)
 
         model = Generator(3, 1, 3)
         model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
diff --git a/modules/control/proc/lineart_anime.py b/modules/control/proc/lineart_anime.py
index 541fcad61..c70da3784 100644
--- a/modules/control/proc/lineart_anime.py
+++ b/modules/control/proc/lineart_anime.py
@@ -117,12 +117,12 @@ class LineartAnimeDetector:
         self.model = model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "netG.pth"
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         norm_layer = functools.partial(nn.InstanceNorm2d, affine=False, track_running_stats=False)
         net = UnetGenerator(3, 1, 8, 64, norm_layer=norm_layer, use_dropout=False)
         ckpt = torch.load(model_path)
diff --git a/modules/control/proc/midas/__init__.py b/modules/control/proc/midas/__init__.py
index ba6d1b0e5..0ca4b4fee 100644
--- a/modules/control/proc/midas/__init__.py
+++ b/modules/control/proc/midas/__init__.py
@@ -17,7 +17,7 @@ class MidasDetector:
         self.model = model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, model_type="dpt_hybrid", filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, model_type="dpt_hybrid", filename=None, cache_dir=None, local_files_only=False):
         if pretrained_model_or_path == "lllyasviel/ControlNet":
             filename = filename or "annotator/ckpts/dpt_hybrid-midas-501f0c75.pt"
         else:
@@ -25,7 +25,7 @@ class MidasDetector:
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         model = MiDaSInference(model_type=model_type, model_path=model_path)
         return cls(model)
 
diff --git a/modules/control/proc/mlsd/__init__.py b/modules/control/proc/mlsd/__init__.py
index ea26c5b0d..e46ac865c 100644
--- a/modules/control/proc/mlsd/__init__.py
+++ b/modules/control/proc/mlsd/__init__.py
@@ -16,7 +16,7 @@ class MLSDdetector:
         self.model = model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None, local_files_only=False):
         if pretrained_model_or_path == "lllyasviel/ControlNet":
             filename = filename or "annotator/ckpts/mlsd_large_512_fp32.pth"
         else:
@@ -24,7 +24,7 @@ class MLSDdetector:
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         model = MobileV2_MLSD_Large()
         model.load_state_dict(torch.load(model_path), strict=True)
         model.eval()
diff --git a/modules/control/proc/normalbae/__init__.py b/modules/control/proc/normalbae/__init__.py
index ba10570c6..16e613dc9 100644
--- a/modules/control/proc/normalbae/__init__.py
+++ b/modules/control/proc/normalbae/__init__.py
@@ -33,12 +33,12 @@ class NormalBaeDetector:
         self.norm = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "scannet.pt"
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         args = types.SimpleNamespace()
         args.mode = 'client'
         args.architecture = 'BN'
diff --git a/modules/control/proc/openpose/__init__.py b/modules/control/proc/openpose/__init__.py
index 80649e213..746351718 100644
--- a/modules/control/proc/openpose/__init__.py
+++ b/modules/control/proc/openpose/__init__.py
@@ -76,7 +76,7 @@ class OpenposeDetector:
         self.face_estimation = face_estimation
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, hand_filename=None, face_filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, hand_filename=None, face_filename=None, cache_dir=None, local_files_only=False):
 
         if pretrained_model_or_path == "lllyasviel/ControlNet":
             filename = filename or "annotator/ckpts/body_pose_model.pth"
@@ -96,9 +96,9 @@ class OpenposeDetector:
             hand_model_path = os.path.join(pretrained_model_or_path, hand_filename)
             face_model_path = os.path.join(face_pretrained_model_or_path, face_filename)
         else:
-            body_model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
-            hand_model_path = hf_hub_download(pretrained_model_or_path, hand_filename, cache_dir=cache_dir)
-            face_model_path = hf_hub_download(face_pretrained_model_or_path, face_filename, cache_dir=cache_dir)
+            body_model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
+            hand_model_path = hf_hub_download(pretrained_model_or_path, hand_filename, cache_dir=cache_dir, local_files_only=local_files_only)
+            face_model_path = hf_hub_download(face_pretrained_model_or_path, face_filename, cache_dir=cache_dir, local_files_only=local_files_only)
 
         body_estimation = Body(body_model_path)
         hand_estimation = Hand(hand_model_path)
diff --git a/modules/control/proc/pidi.py b/modules/control/proc/pidi.py
index 078525f2b..2b18eedf9 100644
--- a/modules/control/proc/pidi.py
+++ b/modules/control/proc/pidi.py
@@ -16,12 +16,12 @@ class PidiNetDetector:
         self.model = model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "table5_pidinet.pth"
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         model = pidinet()
         model.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(model_path)['state_dict'].items()})
         model.eval()
diff --git a/modules/control/proc/segment_anything/__init__.py b/modules/control/proc/segment_anything/__init__.py
index d698b46fc..121421c18 100644
--- a/modules/control/proc/segment_anything/__init__.py
+++ b/modules/control/proc/segment_anything/__init__.py
@@ -23,12 +23,12 @@ class SamDetector:
         self.model = mask_generator
 
     @classmethod
-    def from_pretrained(cls, model_path, filename, model_type, cache_dir=None):
+    def from_pretrained(cls, model_path, filename, model_type, cache_dir=None, local_files_only=False):
         """
         Possible model_type : vit_h, vit_l, vit_b, vit_t
         download weights from https://github.com/facebookresearch/segment-anything
         """
-        model_path = hf_hub_download(model_path, filename, cache_dir=cache_dir)
+        model_path = hf_hub_download(model_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         sam = sam_model_registry[model_type](checkpoint=model_path)
         sam.to(devices.device)
         mask_generator = SamAutomaticMaskGenerator(sam)
diff --git a/modules/control/proc/zoe/__init__.py b/modules/control/proc/zoe/__init__.py
index bb18201e0..a6550d605 100644
--- a/modules/control/proc/zoe/__init__.py
+++ b/modules/control/proc/zoe/__init__.py
@@ -20,12 +20,12 @@ class ZoeDetector:
         self.model = model
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_or_path, model_type="zoedepth", filename=None, cache_dir=None):
+    def from_pretrained(cls, pretrained_model_or_path, model_type="zoedepth", filename=None, cache_dir=None, local_files_only=False):
         filename = filename or "ZoeD_M12_N.pt"
         if os.path.isdir(pretrained_model_or_path):
             model_path = os.path.join(pretrained_model_or_path, filename)
         else:
-            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir)
+            model_path = hf_hub_download(pretrained_model_or_path, filename, cache_dir=cache_dir, local_files_only=local_files_only)
         if model_type == "zoedepth":
             model_cls = ZoeDepth
         elif model_type == "zoedepth_nk":
diff --git a/modules/control/processors.py b/modules/control/processors.py
index baca2fbb3..1f5b0494b 100644
--- a/modules/control/processors.py
+++ b/modules/control/processors.py
@@ -172,15 +172,13 @@ class Processor():
         return f' Processor(id={self.processor_id} model={self.model.__class__.__name__})' if self.processor_id and self.model else ''
 
     def reset(self, processor_id: str = None):
-        from modules.shared import opts
         if self.model is not None:
             debug(f'Control Processor unloaded: id="{self.processor_id}"')
             self.model = None
             self.processor_id = processor_id
             devices.torch_gc(force=True, reason='processor')
-        # self.override = None
-        # devices.torch_gc()
         self.load_config = { 'cache_dir': cache_dir }
+        from modules.shared import opts
         if opts.offline_mode:
             self.load_config["local_files_only"] = True
             os.environ['HF_HUB_OFFLINE'] = '1'
diff --git a/modules/control/units/lite.py b/modules/control/units/lite.py
index fbabe2280..107ebb0a0 100644
--- a/modules/control/units/lite.py
+++ b/modules/control/units/lite.py
@@ -108,8 +108,15 @@ class ControlLLLite():
                     self.model = ControlNetLLLite(model_path)
                 else:
                     import huggingface_hub as hf
+                    offline_config = {}
+                    if opts.offline_mode:
+                        offline_config["local_files_only"] = True
+                        os.environ['HF_HUB_OFFLINE'] = '1'
+                    else:
+                        os.environ.pop('HF_HUB_OFFLINE', None)
+                        os.unsetenv('HF_HUB_OFFLINE')
                     folder, filename = os.path.split(model_path)
-                    model_path = hf.hf_hub_download(repo_id=folder, filename=f'{filename}.safetensors', cache_dir=cache_dir)
+                    model_path = hf.hf_hub_download(repo_id=folder, filename=f'{filename}.safetensors', cache_dir=cache_dir, **offline_config)
                     self.model = ControlNetLLLite(model_path)
                 if self.device is not None:
                     self.model.to(self.device)
diff --git a/modules/control/units/t2iadapter.py b/modules/control/units/t2iadapter.py
index 35ba8ab43..b9e049779 100644
--- a/modules/control/units/t2iadapter.py
+++ b/modules/control/units/t2iadapter.py
@@ -2,7 +2,7 @@ import os
 import time
 from typing import Union
 import threading
-from diffusers import pipelines, StableDiffusionPipeline, StableDiffusionXLPipeline, T2IAdapter, MultiAdapter, StableDiffusionAdapterPipeline, StableDiffusionXLAdapterPipeline # pylint: disable=unused-import
+from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline, T2IAdapter, MultiAdapter, StableDiffusionAdapterPipeline, StableDiffusionXLAdapterPipeline # pylint: disable=unused-import
 from installer import log
 from modules import errors, sd_models
 from modules.control.units import detect
@@ -104,6 +104,13 @@ class Adapter():
                     return
                 model_path, model_args = all_models[model_id]
                 self.load_config.update(model_args)
+                from modules.shared import opts
+                if opts.offline_mode:
+                    self.load_config["local_files_only"] = True
+                    os.environ['HF_HUB_OFFLINE'] = '1'
+                else:
+                    os.environ.pop('HF_HUB_OFFLINE', None)
+                    os.unsetenv('HF_HUB_OFFLINE')
                 if model_path is None:
                     log.error(f'Control {what} model load failed: id="{model_id}" error=unknown model id')
                     return
@@ -168,6 +175,7 @@ class AdapterPipeline():
                 adapter=adapter,
             )
             sd_models.move_model(self.pipeline, pipeline.device)
+            sd_models.apply_balanced_offload(self.pipeline, force=True)
         elif detect.is_sd15(pipeline):
             self.pipeline = StableDiffusionAdapterPipeline(
                 vae=pipeline.vae,
@@ -181,6 +189,7 @@ class AdapterPipeline():
                 adapter=adapter,
             )
             sd_models.move_model(self.pipeline, pipeline.device)
+            sd_models.apply_balanced_offload(self.pipeline, force=True)
         else:
             log.error(f'Control {what} pipeline: class={pipeline.__class__.__name__} unsupported model type')
             return
diff --git a/modules/control/units/xs.py b/modules/control/units/xs.py
index f727a0111..2d56fd7ff 100644
--- a/modules/control/units/xs.py
+++ b/modules/control/units/xs.py
@@ -100,6 +100,12 @@ class ControlNetXS():
                     # log.debug(f'Control {what} model: id="{model_id}" path="{model_path}" already loaded')
                     return
                 self.load_config['time_embedding_mix'] = time_embedding_mix
+                if opts.offline_mode:
+                    self.load_config["local_files_only"] = True
+                    os.environ['HF_HUB_OFFLINE'] = '1'
+                else:
+                    os.environ.pop('HF_HUB_OFFLINE', None)
+                    os.unsetenv('HF_HUB_OFFLINE')
                 log.debug(f'Control {what} model loading: id="{model_id}" path="{model_path}" {self.load_config}')
                 if model_path.endswith('.safetensors'):
                     self.model = ControlNetXSModel.from_single_file(model_path, **self.load_config)
@@ -140,6 +146,7 @@ class ControlNetXSPipeline():
                 controlnet=controlnet, # can be a list
             )
             sd_models.move_model(self.pipeline, pipeline.device)
+            sd_models.apply_balanced_offload(self.pipeline, force=True)
         elif detect.is_sd15(pipeline):
             self.pipeline = StableDiffusionControlNetXSPipeline(
                 vae=pipeline.vae,
@@ -153,6 +160,7 @@ class ControlNetXSPipeline():
                 controlnet=controlnet, # can be a list
             )
             sd_models.move_model(self.pipeline, pipeline.device)
+            sd_models.apply_balanced_offload(self.pipeline, force=True)
         else:
             log.error(f'Control {what} pipeline: class={pipeline.__class__.__name__} unsupported model type')
             return
diff --git a/modules/framepack/framepack_load.py b/modules/framepack/framepack_load.py
index 281bfe010..4b99126ef 100644
--- a/modules/framepack/framepack_load.py
+++ b/modules/framepack/framepack_load.py
@@ -140,14 +140,14 @@ def load_model(variant:str=None, pipeline:str=None, text_encoder:str=None, text_
         shared.log.debug(f'FramePack load: module=llm {model["text_encoder"]}')
         load_args, quant_args = model_quant.get_dit_args({}, module='TE', device_map=True)
         text_encoder = LlamaModel.from_pretrained(model["text_encoder"]["repo"], subfolder=model["text_encoder"]["subfolder"], cache_dir=shared.opts.hfcache_dir, **load_args, **quant_args, **offline_config)
-        tokenizer = LlamaTokenizerFast.from_pretrained(model["tokenizer"]["repo"], subfolder=model["tokenizer"]["subfolder"], cache_dir=shared.opts.hfcache_dir)
+        tokenizer = LlamaTokenizerFast.from_pretrained(model["tokenizer"]["repo"], subfolder=model["tokenizer"]["subfolder"], cache_dir=shared.opts.hfcache_dir, **offline_config)
         text_encoder.requires_grad_(False)
         text_encoder.eval()
         sd_models.move_model(text_encoder, devices.cpu)
 
         shared.log.debug(f'FramePack load: module=te {model["text_encoder_2"]}')
         text_encoder_2 = CLIPTextModel.from_pretrained(model["text_encoder_2"]["repo"], subfolder=model["text_encoder_2"]["subfolder"], torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, **offline_config)
-        tokenizer_2 = CLIPTokenizer.from_pretrained(model["pipeline"]["repo"], subfolder='tokenizer_2', cache_dir=shared.opts.hfcache_dir)
+        tokenizer_2 = CLIPTokenizer.from_pretrained(model["pipeline"]["repo"], subfolder='tokenizer_2', cache_dir=shared.opts.hfcache_dir, **offline_config)
         text_encoder_2.requires_grad_(False)
         text_encoder_2.eval()
         sd_models.move_model(text_encoder_2, devices.cpu)
diff --git a/modules/ipadapter.py b/modules/ipadapter.py
index 88a01e2cd..a74575440 100644
--- a/modules/ipadapter.py
+++ b/modules/ipadapter.py
@@ -190,14 +190,15 @@ def load_image_encoder(pipe: diffusers.DiffusionPipeline, adapter_names: list[st
     if pipe.image_encoder is None or clip_loaded != f'{clip_repo}/{clip_subfolder}':
         jobid = shared.state.begin('Load encoder')
         try:
+            offline_config = { 'local_files_only': True } if shared.opts.offline_mode else {}
             if shared.sd_model_type == 'sd3':
-                image_encoder = transformers.SiglipVisionModel.from_pretrained(clip_repo, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir)
+                image_encoder = transformers.SiglipVisionModel.from_pretrained(clip_repo, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, **offline_config)
             else:
                 if clip_subfolder is None:
-                    image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True)
+                    image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True, **offline_config)
                     shared.log.debug(f'IP adapter load: encoder="{clip_repo}" cls={pipe.image_encoder.__class__.__name__}')
                 else:
-                    image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, subfolder=clip_subfolder, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True)
+                    image_encoder = transformers.CLIPVisionModelWithProjection.from_pretrained(clip_repo, subfolder=clip_subfolder, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, use_safetensors=True, **offline_config)
                     shared.log.debug(f'IP adapter load: encoder="{clip_repo}/{clip_subfolder}" cls={pipe.image_encoder.__class__.__name__}')
             sd_models.clear_caches()
             image_encoder = model_quant.do_post_load_quant(image_encoder, allow=True)
@@ -220,8 +221,9 @@ def load_feature_extractor(pipe):
     if pipe.feature_extractor is None:
         try:
             jobid = shared.state.begin('Load extractor')
+            offline_config = { 'local_files_only': True } if shared.opts.offline_mode else {}
             if shared.sd_model_type == 'sd3':
-                feature_extractor = transformers.SiglipImageProcessor.from_pretrained(SIGLIP_ID, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir)
+                feature_extractor = transformers.SiglipImageProcessor.from_pretrained(SIGLIP_ID, torch_dtype=devices.dtype, cache_dir=shared.opts.hfcache_dir, **offline_config)
             else:
                 feature_extractor = transformers.CLIPImageProcessor()
             if hasattr(pipe, 'register_modules'):
@@ -343,6 +345,8 @@ def apply(pipe, p: processing.StableDiffusionProcessing, adapter_names=[], adapt
             kwargs['weight_name'] = names if len(names) > 1 else names[0]
         if len(revisions) > 0:
             kwargs['revision'] = revisions[0]
+        if shared.opts.offline_mode:
+            kwargs["local_files_only"] = True
         pipe.load_ip_adapter(repos, **kwargs)
         adapters_loaded = names
         if hasattr(p, 'ip_adapter_layers'):
diff --git a/modules/sd_offload.py b/modules/sd_offload.py
index 0d9defd5d..59aea6fd0 100644
--- a/modules/sd_offload.py
+++ b/modules/sd_offload.py
@@ -409,7 +409,7 @@ def report_model_stats(module_name, module):
         shared.log.error(f'Module stats: name={module_name} {e}')
 
 
-def apply_balanced_offload(sd_model=None, exclude=[]):
+def apply_balanced_offload(sd_model=None, exclude=[], force=False):
     global offload_hook_instance # pylint: disable=global-statement
     if shared.opts.diffusers_offload_mode != "balanced":
         return sd_model
@@ -424,7 +424,7 @@ def apply_balanced_offload(sd_model=None, exclude=[]):
         return sd_model
     cached = True
     checkpoint_name = sd_model.sd_checkpoint_info.name if getattr(sd_model, "sd_checkpoint_info", None) is not None else sd_model.__class__.__name__
-    if (offload_hook_instance is None) or (offload_hook_instance.min_watermark != shared.opts.diffusers_offload_min_gpu_memory) or (offload_hook_instance.max_watermark != shared.opts.diffusers_offload_max_gpu_memory) or (checkpoint_name != offload_hook_instance.checkpoint_name):
+    if force or (offload_hook_instance is None) or (offload_hook_instance.min_watermark != shared.opts.diffusers_offload_min_gpu_memory) or (offload_hook_instance.max_watermark != shared.opts.diffusers_offload_max_gpu_memory) or (checkpoint_name != offload_hook_instance.checkpoint_name):
         cached = False
         offload_hook_instance = OffloadHook(checkpoint_name)