diff --git a/extensions-builtin/sdnext-modernui b/extensions-builtin/sdnext-modernui
index e8fa7a839..6871752df 160000
--- a/extensions-builtin/sdnext-modernui
+++ b/extensions-builtin/sdnext-modernui
@@ -1 +1 @@
-Subproject commit e8fa7a839dcce6bf54d8fb06da0469a86f7745bb
+Subproject commit 6871752dfdc373460e5feb13975ec7450df2eec4
diff --git a/modules/loader.py b/modules/loader.py
index 4184840e7..275afadae 100644
--- a/modules/loader.py
+++ b/modules/loader.py
@@ -248,14 +248,14 @@ timer.startup.record("cv2")
 
 class _tqdm_cls:
     def __call__(self, *args, **kwargs):
-        bar_format = 'Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + '{desc}' + '\x1b[0m'
-        return tqdm_lib.tqdm(*args, bar_format=bar_format, ncols=80, colour='#327fba', **kwargs)
+        bar_format = 'Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + '{desc}' + '\x1b[0m'
+        return tqdm_lib.tqdm(*args, bar_format=bar_format, ncols=120, colour='#327fba', **kwargs)
 
 class _tqdm_old(tqdm_lib.tqdm):
     def __init__(self, *args, **kwargs):
         kwargs.pop("name", None)
-        kwargs['bar_format'] = 'Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + '{desc}' + '\x1b[0m'
-        kwargs['ncols'] = 80
+        kwargs['bar_format'] = 'Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + '{desc}' + '\x1b[0m'
+        kwargs['ncols'] = 120
         super().__init__(*args, **kwargs)
 
 try:
diff --git a/modules/logger.py b/modules/logger.py
index 71214a40a..a5c817404 100644
--- a/modules/logger.py
+++ b/modules/logger.py
@@ -291,7 +291,7 @@ def setup_logging(debug=None, trace=None, filename=None):
     if os.environ.get('SD_TRANSFORMERS_DEBUG', None) is not None:
         logging.getLogger("transformers").setLevel(logging.DEBUG)
     else:
-        logging.getLogger("transformers").setLevel(logging.WARNING)
+        logging.getLogger("transformers").setLevel(logging.ERROR)
     if os.environ.get('SD_TORCH_DEBUG', None) is not None:
         logging.getLogger("torch").setLevel(logging.DEBUG)
     else:
diff --git a/modules/video_models/video_modular.py b/modules/modular_load.py
similarity index 57%
rename from modules/video_models/video_modular.py
rename to modules/modular_load.py
index e272cb3e1..76bc28636 100644
--- a/modules/video_models/video_modular.py
+++ b/modules/modular_load.py
@@ -5,7 +5,88 @@ from modules import shared, errors, devices
 from modules.logger import log
 
 
-MIN_LATENT_FRAMES = 7 # decoder floor: fewer latent frames leave the chunked decode with nothing to emit
+class InterruptLogFilter(logging.Filter):
+    """Drops the per-block error dumps the modular runner logs when an interrupt raises through it."""
+    def filter(self, record):
+        return 'Interrupted...' not in record.msg
+
+
+def apply_progress_bar_config(block):
+    kwargs = {
+        "ncols": 120,
+        "colour": "#327fba",
+        "bar_format": "Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} {desc}",
+    }
+    if hasattr(block, "set_progress_bar_config"):
+        block.set_progress_bar_config(**kwargs)
+    for child in getattr(block, "sub_blocks", {}).values():
+        apply_progress_bar_config(child)
+
+
+def install_state_hook(pipe):
+    runner_log = logging.getLogger('diffusers.modular_pipelines.modular_pipeline')
+    if not any(isinstance(f, InterruptLogFilter) for f in runner_log.filters):
+        runner_log.addFilter(InterruptLogFilter())
+
+    def set_phase(phase: str, module: torch.nn.Module | None = None):
+        # every stage runs inside one pipeline call, so the forward hooks are the only place the current stage is visible; state.begin clears the label per job
+        if getattr(pipe, 'sdnext_phase', None) != phase:
+            pipe.sdnext_phase = phase
+            jobid = getattr(pipe, 'sdnext_phaseid', None)
+            shared.state.end(jobid)
+            pipe.sdnext_phaseid = shared.state.begin(phase)
+            log.debug(f'Pipeline: phase={phase} cls={pipe.__class__.__name__} module={module.__class__.__name__ if module is not None else None}')
+
+    def _pre_transformer_hook(module, args): # pylint: disable=unused-argument
+        set_phase('Generate', module)
+        if shared.state.sampling_steps == 0 and getattr(pipe, 'num_timesteps', 0) > 0:
+            shared.state.sampling_steps = pipe.num_timesteps
+        if shared.state.paused:
+            log.debug('Sampling paused')
+            while shared.state.paused:
+                if shared.state.interrupted or shared.state.skipped:
+                    raise AssertionError('Interrupted...')
+                time.sleep(0.1)
+        shared.state.step()
+        if shared.state.interrupted or shared.state.skipped:
+            raise AssertionError('Interrupted...')
+
+    def _pre_text_encode_hook(module, args): # pylint: disable=unused-argument
+        set_phase('TextEncode', module)
+        if shared.state.interrupted or shared.state.skipped:
+            raise AssertionError('Interrupted...')
+
+    def _pre_vae_decode_hook(module, args): # pylint: disable=unused-argument
+        set_phase('Decode', module)
+        if shared.state.interrupted or shared.state.skipped: # fires per tile, so tiled decodes abort promptly
+            raise AssertionError('Interrupted...')
+
+    def _pre_vae_encode_hook(module, args): # pylint: disable=unused-argument
+        set_phase('Encode', module)
+        if shared.state.interrupted or shared.state.skipped: # fires per tile, so tiled encodes abort promptly
+            raise AssertionError('Interrupted...')
+
+    for name in ('transformer', 'transformer_ref'):
+        module = getattr(pipe, name, None)
+        if module is not None:
+            target = getattr(module, 'model', module) # conditioning calls the inner model directly
+            if isinstance(target, torch.nn.Module) and getattr(target, 'sdnext_state_hook', None) is None:
+                target.sdnext_state_hook = target.register_forward_pre_hook(_pre_transformer_hook)
+
+    for name in ('text_encoder', 'text_encoder_2'):
+        module = getattr(pipe, name, None)
+        if module is not None:
+            target = getattr(module, 'model', module) # conditioning calls the inner model directly
+            if isinstance(target, torch.nn.Module) and getattr(target, 'sdnext_state_hook', None) is None:
+                target.sdnext_state_hook = target.register_forward_pre_hook(_pre_text_encode_hook)
+
+    for name in ('vae', 'audio_vae'):
+        decoder = getattr(getattr(pipe, name, None), 'decoder', None) # decode entry points bypass forward, the inner decoder does not
+        if isinstance(decoder, torch.nn.Module) and getattr(decoder, 'sdnext_state_hook', None) is None:
+            decoder.sdnext_state_hook = decoder.register_forward_pre_hook(_pre_vae_decode_hook)
+        encoder = getattr(getattr(pipe, name, None), 'encoder', None) # decode entry points bypass forward, the inner encoder does not
+        if isinstance(encoder, torch.nn.Module) and getattr(encoder, 'sdnext_state_hook', None) is None:
+            encoder.sdnext_state_hook = encoder.register_forward_pre_hook(_pre_vae_encode_hook)
 
 
 def is_modular(obj) -> bool:
@@ -110,157 +191,16 @@ def load_modular_pipe(repo_cls, repo: str, workflow: str | None = None, revision
         missing = missing_components(pipe, workflow)
         pipe.sdnext_missing_components = missing # a caller that can recover a component clears its own entry
         pipe.sdnext_video_workflow = workflow # the workflow this pipe was loaded for, which is what the reference-workflow guard reads; the executed task is chosen per request
-        if hasattr(pipe, 'min_duration') and hasattr(pipe, 'fps'):
-            pipe.sdnext_supported_min_frames = int(pipe.min_duration * pipe.fps) # fresh pipes report the true floor; still mode gates per instance
         log.info(f'Load modular: cls={pipe.__class__.__name__} workflow={workflow} components={loaded} empty={empty} time={time.time()-t0:.2f}')
         if missing:
             # load_components builds each component in its own try/except and reports a failure as a warning on the
             # diffusers logger, so the reason is in the log above this line rather than in the exception path
             log.error(f'Load modular: cls={pipe.__class__.__name__} workflow={workflow} missing={missing} components the workflow requires did not load')
+
+        install_state_hook(pipe)
+        apply_progress_bar_config(pipe._blocks) # pylint: disable=protected-access
         return pipe
     except Exception as e:
         log.error(f'Load modular: repo="{repo}" workflow={workflow} {e}')
         errors.display(e, 'video')
         return None
-
-
-def load_modular(selected, offline_args: dict):
-    return load_modular_pipe(selected.repo_cls, selected.repo, workflow=selected.workflow, revision=selected.repo_revision, offline_args=offline_args, base=selected.base)
-
-
-def apply_minimax_overrides(p, pipe, still: bool = False, audio: bool = True):
-    """Per-generation constraints shared by the video tab and the image path: canvas and frame
-    alignment, the bespoke scheduler guard, tiling, and the audio/still toggles."""
-    if still:
-        audio = False # a sub-second soundtrack is pure waste on a kept single frame
-    multiple = pipe.canvas_multiple
-    p.task_args['width'] = multiple * (p.width // multiple)
-    p.task_args['height'] = multiple * (p.height // multiple)
-    set_still(pipe, still)
-    if still:
-        frames = 5 # two latent frames; decode pads to the decoder floor and only the first frame is kept
-        log.info(f'Pipeline: cls={pipe.__class__.__name__} mode=still')
-    else:
-        frames = max(getattr(p, 'frames', 1), getattr(pipe, 'sdnext_supported_min_frames', 120))
-        while frames % pipe.vae_frames_per_chunk != pipe.vae_latents_per_chunk: # frame counts align to 17n+5
-            frames += 1
-        max_frames = int(pipe.max_duration * pipe.fps)
-        while frames > max_frames:
-            frames -= pipe.vae_frames_per_chunk
-    if frames != getattr(p, 'frames', None):
-        log.debug(f'Pipeline: cls={pipe.__class__.__name__} frames={getattr(p, "frames", None)} aligned={frames}')
-    p.frames = frames
-    p.task_args['num_frames'] = frames
-    p.steps = max(2, p.steps)
-    p.task_args['num_inference_steps'] = p.steps
-    pipe.num_timesteps = p.steps - 1 # sigma grid includes the terminal point; feeds the progress total
-    if p.sampler_name not in ('None', 'Default'):
-        log.warning(f'Pipeline: cls={pipe.__class__.__name__} sampler={p.sampler_name} unsupported: using model default')
-    p.sampler_name = 'Default' # the model default is the bespoke scheduler pair, which discrete samplers must not replace
-    pipe.vae.enable_tiling() # model always tiles; the shared vae params path may have disabled it
-    set_audio(pipe, audio)
-    p.task_args['output'] = ['videos', 'audio', 'sampling_rate'] if audio else ['videos']
-    p.task_args['output_type'] = 'pil' # the image path otherwise requests latent output, which the decode block rejects
-    p.video_still = still
-
-
-def set_still(pipe, enabled: bool = True):
-    """Toggle sub-floor generation for single-frame output. The duration floor is lifted only
-    while the instance flag is set, so other pipes of the class and later normal runs keep the
-    supported floor; decoded latents below the decoder floor are padded by duplicating the
-    trailing latent. The causal VAE keeps padding out of frame 0."""
-    cls = type(pipe)
-    if getattr(cls, 'sdnext_min_duration_orig', None) is None:
-        orig = cls.min_duration
-        cls.sdnext_min_duration_orig = orig
-        cls.min_duration = property(lambda self: 0.0 if getattr(self, 'sdnext_still_mode', False) else orig.fget(self))
-    pipe.sdnext_still_mode = enabled
-    if not enabled:
-        return
-    vae = getattr(pipe, 'vae', None)
-    if vae is not None and getattr(vae, 'sdnext_orig_decode', None) is None:
-        vae.sdnext_orig_decode = vae.decode
-        def padded_decode(z, *args, **kwargs):
-            if z.ndim == 5 and z.shape[2] < MIN_LATENT_FRAMES:
-                pad = z[:, :, -1:].repeat(1, 1, MIN_LATENT_FRAMES - z.shape[2], 1, 1)
-                z = torch.cat([z, pad], dim=2)
-            return vae.sdnext_orig_decode(z, *args, **kwargs)
-        vae.decode = padded_decode
-
-
-def set_audio(pipe, enabled: bool):
-    """Pop or restore the audio decode block. The joint denoise still carries the audio rows
-    (a few percent of the sequence), but without the block the audio VAE never runs.
-    Operates on the backing block tree: the public blocks property deep-copies per access."""
-    blocks = getattr(pipe, '_blocks', None) # pylint: disable=protected-access
-    decode = blocks.sub_blocks.get('decode', None) if blocks is not None and hasattr(blocks, 'sub_blocks') else None
-    sub = getattr(decode, 'sub_blocks', None)
-    if sub is None:
-        return
-    if enabled and 'audio' not in sub:
-        stashed = getattr(pipe, 'sdnext_audio_decode_block', None)
-        if stashed is not None:
-            sub.insert('audio', stashed, len(sub))
-            log.debug(f'Pipeline: cls={pipe.__class__.__name__} audio=enabled')
-    elif not enabled and 'audio' in sub:
-        pipe.sdnext_audio_decode_block = sub.pop('audio')
-        log.debug(f'Pipeline: cls={pipe.__class__.__name__} audio=disabled')
-
-
-class InterruptLogFilter(logging.Filter):
-    """Drops the per-block error dumps the modular runner logs when an interrupt raises through it."""
-    def filter(self, record):
-        return 'Interrupted...' not in record.getMessage()
-
-
-def install_state_hook(pipe):
-    runner_log = logging.getLogger('diffusers.modular_pipelines.modular_pipeline')
-    if not any(isinstance(f, InterruptLogFilter) for f in runner_log.filters):
-        runner_log.addFilter(InterruptLogFilter())
-
-    def set_phase(phase: str):
-        # every stage runs inside one pipeline call, so the forward hooks are the only
-        # place the current stage is visible; state.begin clears the label per job
-        if getattr(pipe, 'sdnext_phase', None) != phase:
-            pipe.sdnext_phase = phase
-            shared.state.textinfo = phase
-            log.debug(f'Pipeline: cls={pipe.__class__.__name__} phase={phase}')
-
-    def state_hook(module, args): # pylint: disable=unused-argument
-        set_phase('Generate')
-        if shared.state.sampling_steps == 0 and getattr(pipe, 'num_timesteps', 0) > 0:
-            shared.state.sampling_steps = pipe.num_timesteps
-        if shared.state.paused:
-            log.debug('Sampling paused')
-            while shared.state.paused:
-                if shared.state.interrupted or shared.state.skipped:
-                    raise AssertionError('Interrupted...')
-                time.sleep(0.1)
-        shared.state.step()
-        if shared.state.interrupted or shared.state.skipped:
-            raise AssertionError('Interrupted...')
-
-    def encode_hook(module, args): # pylint: disable=unused-argument
-        set_phase('TextEncode')
-        if shared.state.interrupted or shared.state.skipped:
-            raise AssertionError('Interrupted...')
-
-    def decode_hook(module, args): # pylint: disable=unused-argument
-        set_phase('Decode')
-        if shared.state.interrupted or shared.state.skipped: # fires per tile, so tiled decodes abort promptly
-            raise AssertionError('Interrupted...')
-
-    for name in ('transformer', 'transformer_ref'):
-        module = getattr(pipe, name, None)
-        if module is None or getattr(module, 'sdnext_state_hook', None) is not None:
-            continue
-        module.sdnext_state_hook = module.register_forward_pre_hook(state_hook)
-    text_encoder = getattr(pipe, 'text_encoder', None)
-    if text_encoder is not None:
-        target = getattr(text_encoder, 'model', text_encoder) # conditioning calls the inner model directly
-        if isinstance(target, torch.nn.Module) and getattr(target, 'sdnext_state_hook', None) is None:
-            target.sdnext_state_hook = target.register_forward_pre_hook(encode_hook)
-    for name in ('vae', 'audio_vae'):
-        decoder = getattr(getattr(pipe, name, None), 'decoder', None) # decode entry points bypass forward, the inner decoder does not
-        if isinstance(decoder, torch.nn.Module) and getattr(decoder, 'sdnext_state_hook', None) is None:
-            decoder.sdnext_state_hook = decoder.register_forward_pre_hook(decode_hook)
diff --git a/modules/postprocess/sdupscaler_model.py b/modules/postprocess/sdupscaler_model.py
index a13cdf30d..01a6efc73 100644
--- a/modules/postprocess/sdupscaler_model.py
+++ b/modules/postprocess/sdupscaler_model.py
@@ -33,7 +33,7 @@ class UpscalerDiffusion(Upscaler):
         else:
             model = diffusers.DiffusionPipeline.from_pretrained(scaler.data_path, cache_dir=shared.opts.diffusers_dir, torch_dtype=devices.dtype)
             if hasattr(model, "set_progress_bar_config"):
-                model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + 'Upscale', ncols=80, colour='#327fba')
+                model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + 'Upscale', ncols=120, colour='#327fba')
             set_diffuser_options(model, vae=None, op='upscaler')
             self.models[path] = model
         return self.models[path]
diff --git a/modules/processing.py b/modules/processing.py
index f897dae82..2134a8ac5 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -428,7 +428,8 @@ def print_stats():
 
         from modules.sd_models_compile import update_compile_times
         update_compile_times()
-        dynamo_dct = timer.dynamo.dct(min_time=2.0, no_total=True)
+        dynamo_dct = timer.dynamo.dct(min_time=0.5, no_total=True)
+        timer.dynamo.reset()
         if dynamo_dct:
             log.debug(f'Processed: dynamo={dynamo_dct}')
 
diff --git a/modules/processing_args.py b/modules/processing_args.py
index 0755f169a..d6f427131 100644
--- a/modules/processing_args.py
+++ b/modules/processing_args.py
@@ -183,8 +183,21 @@ def get_defaults(model, kwargs):
     remove = ['return_dict', 'output_type', 'num_images_per_prompt', 'callback', 'callback_on_step_end_tensor_inputs']
     default_cfg = 0
     try:
-        signature = inspect.signature(type(model).__call__, follow_wrapped=True)
-        defaults = {k: v.default for k, v in signature.parameters.items() if v.default is not inspect.Parameter.empty and v.default is not None} # get all defaults
+        defaults = {}
+        if hasattr(model, 'blocks') and hasattr(model.blocks, 'inputs'):
+            for input_param in model.blocks.inputs:
+                if input_param.name is None:
+                    continue
+                if input_param.default is None:
+                    continue
+                if input_param.name in kwargs or input_param.name in remove:
+                    continue
+                defaults[input_param.name] = input_param.default
+
+        if not defaults:
+            signature = inspect.signature(type(model).__call__, follow_wrapped=True)
+            defaults = {k: v.default for k, v in signature.parameters.items() if v.default is not inspect.Parameter.empty and v.default is not None} # get all defaults
+
         defaults = {k: v for k, v in defaults.items() if k not in kwargs} # only log defaults that are not already set by kwargs
         defaults = {k: v for k, v in defaults.items() if k not in remove} # remove common args that are not useful to log
         log.debug(f'Pipeline: cls={model.__class__.__name__} defaults={defaults}')
@@ -222,9 +235,9 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:l
             model.register_to_config(boundary_ratio=boundary_target)
     if hasattr(model, "set_progress_bar_config"):
         if disable_pbar:
-            model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + desc, ncols=80, colour='#327fba', disable=disable_pbar)
+            model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + desc, ncols=120, colour='#327fba', disable=disable_pbar)
         else:
-            model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + desc, ncols=80, colour='#327fba')
+            model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + desc, ncols=120, colour='#327fba')
 
     possible = get_params(model)
 
diff --git a/modules/processing_diffusers.py b/modules/processing_diffusers.py
index 46b7d884b..d7361fc87 100644
--- a/modules/processing_diffusers.py
+++ b/modules/processing_diffusers.py
@@ -76,6 +76,7 @@ def process_pre(p: processing.StableDiffusionProcessing):
     log.info('Processing modifiers: apply')
     try:
         # apply-with-unapply
+        # sd_hijack_compile.install()
         sd_models_compile.check_deepcache(enable=True)
         ipadapter.apply(shared.sd_model, p)
         token_merge.apply_token_merging(shared.sd_model)
@@ -235,6 +236,9 @@ def process_base(p: processing.StableDiffusionProcessing):
     if hasattr(shared.sd_model, 'postprocess') and callable(shared.sd_model.postprocess):
         output = shared.sd_model.postprocess(p, output)
 
+    if hasattr(shared.sd_model, 'sdnext_phaseid'):
+        shared.state.end(shared.sd_model.sdnext_phaseid)
+        shared.sd_model.sdnext_phaseid = None
     shared.state.end(jobid)
     shared.state.nextjob()
     return output
@@ -530,8 +534,8 @@ def update_pipeline(sd_model, p: processing.StableDiffusionProcessing):
     updated_model = sd_model
     if 'MiniMaxH3' in sd_model.__class__.__name__ and not isinstance(p, processing.StableDiffusionProcessingVideo):
         # image tabs run the model in still mode; the video tab applies its own overrides
-        from modules.video_models import video_modular
-        video_modular.apply_minimax_overrides(p, sd_model, still=True, audio=False)
+        from modules.video_models import video_minimax
+        video_minimax.apply_overrides(p, sd_model, still=True, audio=False)
         if getattr(p, 'detailer_enabled', False):
             log.warning(f'Processing: cls={sd_model.__class__.__name__} detailer not supported')
             p.detailer_enabled = False
diff --git a/modules/sd_hijack_compile.py b/modules/sd_hijack_compile.py
new file mode 100644
index 000000000..1b306c9ec
--- /dev/null
+++ b/modules/sd_hijack_compile.py
@@ -0,0 +1,31 @@
+import time
+import logging
+from modules.timer import dynamo
+
+
+fn = None
+ts = None
+
+
+class CompilationLogInterceptor(logging.Handler):
+    def emit(self, record):
+        try:
+            global fn, ts # pylint: disable=global-statement
+            if 'torchdynamo start tracing' in record.msg:
+                fn = record.msg.split('torchdynamo start tracing')[-1].strip() # extract first string after 'torchdynamo start tracing' and start timer
+                fn = fn.split(' ')[0] # extract first word after 'torchdynamo start tracing'
+                ts = time.time()
+            if 'run_gc_after_compile' in record.msg:
+                if fn is not None:
+                    dynamo.ts(fn, ts) # log the time taken for compilation
+                    fn = None
+        except Exception:
+            pass
+
+
+def install():
+    dynamo_logger = logging.getLogger("torch._dynamo")
+    dynamo_logger.setLevel(logging.INFO)
+    if not any(isinstance(h, CompilationLogInterceptor) for h in dynamo_logger.handlers):
+        dynamo_interceptor = CompilationLogInterceptor()
+        dynamo_logger.addHandler(dynamo_interceptor)
diff --git a/modules/sd_hijack_hfhub.py b/modules/sd_hijack_hfhub.py
index 42f963ee3..87c62586d 100644
--- a/modules/sd_hijack_hfhub.py
+++ b/modules/sd_hijack_hfhub.py
@@ -1,5 +1,6 @@
 import os
 import time
+import tqdm
 from modules.logger import log
 
 
@@ -54,9 +55,10 @@ def xet_get_hijack(*args, **kwargs):
     if fn and not fn.endswith(".json"):
         log.debug(f'Download: type=xet mode="{opts.hf_transfer_mode}" fn="{fn}" size={size}')
     debug(f'Download start: type=xet args={args} kwargs={kwargs}')
-    # import tqdm # TODO xet_download: hijack progress bar
-    # bar_format = 'Download {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m' + '{desc}' + '\x1b[0m'
-    # kwargs['_tqdm_bar'] = tqdm.tqdm(*args, bar_format=bar_format, ncols=80, colour='#327fba', **kwargs)
+
+    bar_format = 'Download {percentage:3.0f}% {bar:15} {n:.1f}/{total:.1f}{postfix} {elapsed} {remaining} ' + '\x1b[38;5;71m' + '{desc}' + '\x1b[0m'
+    kwargs['_tqdm_bar'] = tqdm.tqdm(*args, bar_format=bar_format, colour='#327fba', unit='MiB', unit_scale=1/(1024*1024), desc=fn, total=size, ncols=120)
+
     res = orig_xet_get(*args, **kwargs)
     debug(f'Download end: type=xet res={res}')
     state.end(jobid)
diff --git a/modules/sd_models.py b/modules/sd_models.py
index daa4771ae..a393652fd 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -893,7 +893,7 @@ def set_defaults(sd_model, checkpoint_info: CheckpointInfo):
     sd_model.logvar = sd_model.logvar.to(devices.device) if hasattr(sd_model, 'logvar') else None # fix for training
     shared.opts.data["sd_checkpoint_hash"] = checkpoint_info.sha256
     if hasattr(sd_model, "set_progress_bar_config"):
-        sd_model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining}', ncols=80, colour='#327fba')
+        sd_model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining}', ncols=120, colour='#327fba')
 
 
 def load_diffuser(checkpoint_info: CheckpointInfo | None = None, op='model', revision=None): # pylint: disable=unused-argument
diff --git a/modules/sd_models_compile.py b/modules/sd_models_compile.py
index d49cc8c48..a6c4acf54 100644
--- a/modules/sd_models_compile.py
+++ b/modules/sd_models_compile.py
@@ -391,17 +391,6 @@ def update_compile_times():
         try:
             times = [float(t.strip()) for t in parts[1:] if t.strip()]
             if times:
-                # parsed.append((fn, sum(times), len(times), max(times)))
                 dynamo.add(fn, round(sum(times), 2))
         except ValueError:
             continue
-    """
-    parsed.sort(key=lambda x: x[1], reverse=True)
-    results = {}
-    min_time = 0.1
-    for fn, total, count, max_val in parsed:
-        if total > min_time:
-            dynamo.ts(fn, total)
-            results[fn] = { "total": round(total, 2), "count": count, "avg": round(total / count, 2), "max": round(max_val, 2) }
-    return results
-    """
diff --git a/modules/sd_offload.py b/modules/sd_offload.py
index 543d5826a..14defe646 100644
--- a/modules/sd_offload.py
+++ b/modules/sd_offload.py
@@ -164,7 +164,7 @@ def apply_group_offload_component(module, module_name: str, main: bool) -> bool:
         module = accelerate.hooks.remove_hook_from_module(module, recurse=True)
     remove_group_offload_component(module)
     module.requires_grad_(False)
-    log.debug(f'Offload: type=group op=apply module={module_name} pin={cfg["use_stream"] and not cfg["low_cpu_mem_usage"]}') # before the apply: pinning large components takes a while and would otherwise run silently
+    log.debug(f'Offload: type=group op=apply type={shared.opts.group_offload_type} module={module_name} pin={cfg["use_stream"] and not cfg["low_cpu_mem_usage"]}') # before the apply: pinning large components takes a while and would otherwise run silently
     apply_group_offloading(module, onload_device=devices.device, offload_device=devices.cpu, **cfg)
     module.sdnext_group_offload_sig = sig
     return True
diff --git a/modules/shared_defaults.py b/modules/shared_defaults.py
index 512f5913b..c18a59e24 100644
--- a/modules/shared_defaults.py
+++ b/modules/shared_defaults.py
@@ -21,12 +21,12 @@ def get_default_modes(cmd_opts, mem_stat):
                 cmd_opts.medvram = True # VAE Tiling and other stuff
                 default_offload_mode = "balanced"
                 default_diffusers_offload_min_gpu_memory = 0
-                default_diffusers_offload_always = ', '.join(['T5EncoderModel', 'UMT5EncoderModel'])
+                default_diffusers_offload_always = ''
                 log.info(f"Device detect: memory={gpu_memory:.1f} default=balanced optimization=medvram")
-            elif gpu_memory >= 24:
+            elif gpu_memory >= 22:
                 default_offload_mode = "balanced"
                 default_diffusers_offload_max_gpu_memory = 0.8
-                default_diffusers_offload_always = ', '.join(['T5EncoderModel', 'UMT5EncoderModel'])
+                default_diffusers_offload_always = ''
                 default_diffusers_offload_never = ', '.join(['CLIPTextModel', 'CLIPTextModelWithProjection', 'AutoencoderKL'])
                 log.info(f"Device detect: memory={gpu_memory:.1f} default=balanced optimization=highvram")
             else:
diff --git a/modules/video_models/video_load.py b/modules/video_models/video_load.py
index ad3fc2237..7b68e3231 100644
--- a/modules/video_models/video_load.py
+++ b/modules/video_models/video_load.py
@@ -4,9 +4,9 @@ import copy
 import time
 import transformers
 import diffusers
-from modules import shared, errors, sd_models, sd_checkpoint, model_quant, devices, sd_hijack_te, sd_hijack_vae
+from modules import shared, errors, sd_models, sd_checkpoint, model_quant, devices, sd_hijack_te, sd_hijack_vae, modular_load
 from modules.logger import log
-from modules.video_models import models_def, video_utils, video_overrides, video_cache, video_modular
+from modules.video_models import models_def, video_utils, video_overrides, video_cache
 
 
 def _loader(component):
@@ -151,8 +151,9 @@ def load_model(selected: models_def.Model):
 
     # model
     try:
-        if selected.workflow is not None or video_modular.is_modular(selected.repo_cls):
-            shared.sd_model = video_modular.load_modular(selected, offline_args)
+        if selected.workflow is not None or modular_load.is_modular(selected.repo_cls):
+            from modules.modular_load import load_modular_pipe
+            return load_modular_pipe(selected.repo_cls, selected.repo, workflow=selected.workflow, revision=selected.repo_revision, offline_args=offline_args, base=selected.base)
         elif selected.repo_cls is None:
             shared.sd_model = load_custom(selected.repo)
         else:
@@ -208,12 +209,12 @@ def load_model(selected: models_def.Model):
         shared.sd_model.vae.enable_tiling()
         tiling = True
     if hasattr(shared.sd_model, "set_progress_bar_config"):
-        shared.sd_model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m', ncols=80, colour='#327fba')
+        shared.sd_model.set_progress_bar_config(bar_format='Progress {rate_fmt}{postfix} {bar:15} {percentage:3.0f}% {n_fmt}/{total_fmt} {elapsed} {remaining} ' + '\x1b[38;5;71m', ncols=120, colour='#327fba')
 
     shared.sd_model = model_quant.do_post_load_quant(shared.sd_model, allow=False)
     sd_models.set_diffuser_offload(shared.sd_model)
-    if video_modular.is_modular(shared.sd_model):
-        video_modular.install_state_hook(shared.sd_model)
+    if modular_load.is_modular(shared.sd_model):
+        modular_load.install_state_hook(shared.sd_model)
 
     loaded_model = selected.name
     msg = f'Load video: cls={shared.sd_model.__class__.__name__} model="{selected.name}" time={t1-t0:.2f}'
diff --git a/modules/video_models/video_minimax.py b/modules/video_models/video_minimax.py
new file mode 100644
index 000000000..31e6fa34a
--- /dev/null
+++ b/modules/video_models/video_minimax.py
@@ -0,0 +1,84 @@
+import torch
+from modules.logger import log
+
+
+MIN_LATENT_FRAMES = 7 # decoder floor: fewer latent frames leave the chunked decode with nothing to emit
+
+
+def apply_overrides(p, pipe, still: bool = False, audio: bool = True):
+    """Per-generation constraints shared by the video tab and the image path: canvas and frame
+    alignment, the bespoke scheduler guard, tiling, and the audio/still toggles."""
+    if still:
+        audio = False # a sub-second soundtrack is pure waste on a kept single frame
+    multiple = pipe.canvas_multiple
+    p.task_args['width'] = multiple * (p.width // multiple)
+    p.task_args['height'] = multiple * (p.height // multiple)
+    set_still(pipe, still)
+    if still:
+        frames = 5 # two latent frames; decode pads to the decoder floor and only the first frame is kept
+        log.info(f'Pipeline: cls={pipe.__class__.__name__} mode=still')
+    else:
+        frames = max(getattr(p, 'frames', 1), getattr(pipe, 'sdnext_supported_min_frames', 120))
+        while frames % pipe.vae_frames_per_chunk != pipe.vae_latents_per_chunk: # frame counts align to 17n+5
+            frames += 1
+        max_frames = int(pipe.max_duration * pipe.fps)
+        while frames > max_frames:
+            frames -= pipe.vae_frames_per_chunk
+    if frames != getattr(p, 'frames', None):
+        log.debug(f'Pipeline: cls={pipe.__class__.__name__} frames={getattr(p, "frames", None)} aligned={frames}')
+    p.frames = frames
+    p.task_args['num_frames'] = frames
+    p.steps = max(2, p.steps)
+    p.task_args['num_inference_steps'] = p.steps
+    pipe.num_timesteps = p.steps - 1 # sigma grid includes the terminal point; feeds the progress total
+    if p.sampler_name not in ('None', 'Default'):
+        log.warning(f'Pipeline: cls={pipe.__class__.__name__} sampler={p.sampler_name} unsupported: using model default')
+    p.sampler_name = 'Default' # the model default is the bespoke scheduler pair, which discrete samplers must not replace
+    pipe.vae.enable_tiling() # model always tiles; the shared vae params path may have disabled it
+    set_audio(pipe, audio)
+    p.task_args['output'] = ['videos', 'audio', 'sampling_rate'] if audio else ['videos']
+    p.task_args['output_type'] = 'pil' # the image path otherwise requests latent output, which the decode block rejects
+    p.video_still = still
+
+
+def set_still(pipe, enabled: bool = True):
+    """Toggle sub-floor generation for single-frame output. The duration floor is lifted only
+    while the instance flag is set, so other pipes of the class and later normal runs keep the
+    supported floor; decoded latents below the decoder floor are padded by duplicating the
+    trailing latent. The causal VAE keeps padding out of frame 0."""
+    cls = type(pipe)
+    if getattr(cls, 'sdnext_min_duration_orig', None) is None:
+        orig = cls.min_duration
+        cls.sdnext_min_duration_orig = orig
+        cls.min_duration = property(lambda self: 0.0 if getattr(self, 'sdnext_still_mode', False) else orig.fget(self))
+    pipe.sdnext_still_mode = enabled
+    if not enabled:
+        return
+    vae = getattr(pipe, 'vae', None)
+    if vae is not None and getattr(vae, 'sdnext_orig_decode', None) is None:
+        vae.sdnext_orig_decode = vae.decode
+        def padded_decode(z, *args, **kwargs):
+            if z.ndim == 5 and z.shape[2] < MIN_LATENT_FRAMES:
+                pad = z[:, :, -1:].repeat(1, 1, MIN_LATENT_FRAMES - z.shape[2], 1, 1)
+                z = torch.cat([z, pad], dim=2)
+            return vae.sdnext_orig_decode(z, *args, **kwargs)
+        vae.decode = padded_decode
+
+
+def set_audio(pipe, enabled: bool):
+    """Pop or restore the audio decode block. The joint denoise still carries the audio rows
+    (a few percent of the sequence), but without the block the audio VAE never runs.
+    Operates on the backing block tree: the public blocks property deep-copies per access."""
+    blocks = getattr(pipe, '_blocks', None) # pylint: disable=protected-access
+    decode = blocks.sub_blocks.get('decode', None) if blocks is not None and hasattr(blocks, 'sub_blocks') else None
+    sub = getattr(decode, 'sub_blocks', None)
+    if sub is None:
+        return
+    if enabled and 'audio' not in sub:
+        stashed = getattr(pipe, 'sdnext_audio_decode_block', None)
+        if stashed is not None:
+            sub.insert('audio', stashed, len(sub))
+            log.debug(f'Pipeline: cls={pipe.__class__.__name__} audio=enabled')
+    elif not enabled and 'audio' in sub:
+        pipe.sdnext_audio_decode_block = sub.pop('audio')
+        log.debug(f'Pipeline: cls={pipe.__class__.__name__} audio=disabled')
diff --git a/modules/video_models/video_overrides.py b/modules/video_models/video_overrides.py
index 2c9b8c79e..a294afd4d 100644
--- a/modules/video_models/video_overrides.py
+++ b/modules/video_models/video_overrides.py
@@ -4,7 +4,6 @@ import diffusers
 from modules import shared, processing, devices
 from modules.logger import log
 from modules.video_models.models_def import Model
-from modules.video_models import video_modular
 
 
 debug = log.trace if os.environ.get('SD_VIDEO_DEBUG', None) is not None else lambda *args, **kwargs: None
@@ -107,4 +106,5 @@ def set_overrides(p: processing.StableDiffusionProcessingVideo, selected: Model)
         shared.sd_model.transformer.set_attention_backend("flex")
     # MiniMax H3
     if 'MiniMaxH3' in cls:
-        video_modular.apply_minimax_overrides(p, shared.sd_model, still=getattr(p, 'video_still', False), audio=getattr(p, 'video_audio', True))
+        from modules.video_models import video_minimax
+        video_minimax.apply_overrides(p, shared.sd_model, still=getattr(p, 'video_still', False), audio=getattr(p, 'video_audio', True))
diff --git a/modules/video_models/video_run.py b/modules/video_models/video_run.py
index 92296b4b5..6444607eb 100644
--- a/modules/video_models/video_run.py
+++ b/modules/video_models/video_run.py
@@ -2,9 +2,9 @@ import os
 import copy
 import time
 from dataclasses import dataclass
-from modules import shared, errors, sd_models, processing, devices, images, ui_common, scripts_manager
+from modules import shared, errors, sd_models, processing, devices, images, ui_common, scripts_manager, modular_load
 from modules.logger import log
-from modules.video_models import models_def, video_utils, video_load, video_vae, video_overrides, video_save, video_modular
+from modules.video_models import models_def, video_utils, video_load, video_vae, video_overrides, video_save
 from modules.paths import resolve_output_path
 
 
@@ -49,7 +49,7 @@ def resolve_model(engine: str | None, model: str | None) -> tuple[models_def.Mod
         raise VideoError(f'no video model loaded: cls={cls} select engine and model or load a video-capable checkpoint first', 400)
     pipe = shared.sd_model
     workflow = getattr(pipe, 'sdnext_video_workflow', None)
-    if workflow is None and video_modular.is_modular(pipe):
+    if workflow is None and modular_load.is_modular(pipe):
         workflow = models_def.workflow_for_class(cls) or 'auto' # modular pipes dispatch on inputs, so any workflow marker selects the modular branch
     ckpt = getattr(pipe, 'sd_checkpoint_info', None)
     selected = models_def.Model(
diff --git a/pipelines/generic_text_encoder.py b/pipelines/generic_text_encoder.py
index a67ac308d..c1b6a57a2 100644
--- a/pipelines/generic_text_encoder.py
+++ b/pipelines/generic_text_encoder.py
@@ -11,7 +11,7 @@ from pipelines.generic_shared import shared_te_map
 debug = os.environ.get('SD_LOAD_DEBUG', None) is not None
 
 
-def get_shared(cls, repo_id, subfolder=None, variant=None):
+def get_shared(cls, repo_id, subfolder=None, variant=None, shared_id: str | None = None):
     args = {}
     if variant is not None:
         args['variant'] = variant
@@ -22,7 +22,8 @@ def get_shared(cls, repo_id, subfolder=None, variant=None):
         if isinstance(identifiers, str):
             identifiers = [identifiers]
         identifiers = [identifier.lower() for identifier in identifiers if identifier is not None]
-        if item['cls'] == cls and (not identifiers or any(identifier in repo_id.lower() for identifier in identifiers)):
+        shared_id = shared_id or repo_id.lower()
+        if item['cls'] == cls and (not identifiers or any(identifier in shared_id for identifier in identifiers)):
             if item.get('config_class', None) is not None and item.get('config_path', None) is not None:
                 with open(item['config_path'], encoding='utf8') as f:
                     args['config'] = item['config_class'](**json.load(f))
@@ -91,6 +92,7 @@ def load_text_encoder(
         modules_to_not_convert=None,
         modules_dtype_dict=None,
         use_safetensors=True,
+        shared_id: str | None = None,
         **kwargs):
 
     if shared.state.interrupted:
@@ -146,7 +148,7 @@ def load_text_encoder(
         # 3. load shared from repo
         if allow_shared and (text_encoder is None):
             log.debug(f'Load model: text_encoder="{repo_id}" cls={cls_name.__name__} quant="{quant_type}" loader={get_loader("transformers")}')
-            target_repo, extra_args = get_shared(cls_name, repo_id, subfolder=subfolder, variant=variant)
+            target_repo, extra_args = get_shared(cls_name, repo_id, subfolder=subfolder, variant=variant, shared_id=shared_id)
             text_encoder = cls_name.from_pretrained(
                 target_repo,
                 cache_dir=shared.opts.hfcache_dir,
diff --git a/pipelines/model_minimax.py b/pipelines/model_minimax.py
index 0d13cd065..4c7ba4f94 100644
--- a/pipelines/model_minimax.py
+++ b/pipelines/model_minimax.py
@@ -4,7 +4,8 @@ from modules.logger import log
 
 
 def load_minimax(checkpoint_info, diffusers_load_config=None): # pylint: disable=unused-argument
-    from modules.video_models import video_modular, video_load
+    from modules.video_models import video_load
+    from modules.modular_load import load_modular_pipe
     repo_id = sd_models.path_to_repo(checkpoint_info)
     sd_models.hf_auth_check(checkpoint_info)
     if repo_id is None or repo_id.lower() == 'none':
@@ -14,7 +15,7 @@ def load_minimax(checkpoint_info, diffusers_load_config=None): # pylint: disable
     log.debug(f'Load model: type=MiniMaxH3 repo="{repo_id}" workflow={workflow} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype}')
 
     repo_cls = diffusers.MiniMaxH3ModularPipeline
-    pipe = video_modular.load_modular_pipe(
+    pipe = load_modular_pipe(
         repo_cls,
         repo_id,
         workflow=workflow,
@@ -24,13 +25,10 @@ def load_minimax(checkpoint_info, diffusers_load_config=None): # pylint: disable
     )
     if pipe is None:
         return None
-    missing = video_modular.missing_components(pipe, workflow)
-    if missing:
-        # a component that failed to build is unusable, and loading it by another route only defers the failure into generation as corrupt output
-        log.error(f'Load model: type=MiniMaxH3 repo="{repo_id}" workflow={workflow} missing={missing}')
-        return None
 
-    video_modular.install_state_hook(pipe)
+    if hasattr(pipe, 'min_duration') and hasattr(pipe, 'fps'):
+        pipe.sdnext_supported_min_frames = int(pipe.min_duration * pipe.fps) # fresh pipes report the true floor; still mode gates per instance
+
     video_load.loaded_model = None # image-path load invalidates the video tab's name cache
     if hasattr(pipe, 'vae') and hasattr(pipe.vae, 'enable_tiling'):
         pipe.vae.enable_tiling()
diff --git a/scripts/pulid/eva_clip/pretrained.py b/scripts/pulid/eva_clip/pretrained.py
index bf957db74..f7a1afe4c 100644
--- a/scripts/pulid/eva_clip/pretrained.py
+++ b/scripts/pulid/eva_clip/pretrained.py
@@ -266,7 +266,7 @@ def download_pretrained_from_url(
             return download_target
 
     with urllib.request.urlopen(url) as source, open(download_target, "wb") as output:
-        with tqdm(total=int(source.headers.get("Content-Length")), ncols=80, unit='iB', unit_scale=True) as loop:
+        with tqdm(total=int(source.headers.get("Content-Length")), ncols=120, unit='iB', unit_scale=True) as loop:
             while True:
                 buffer = source.read(8192)
                 if not buffer:
diff --git a/scripts/pulid/pulid_sdxl.py b/scripts/pulid/pulid_sdxl.py
index 74eca27b8..a36283be3 100644
--- a/scripts/pulid/pulid_sdxl.py
+++ b/scripts/pulid/pulid_sdxl.py
@@ -283,7 +283,7 @@ class StableDiffusionXLPuLIDPipeline:
         debug(f'PulID embedding: cond={id_embedding.shape} uncond={uncond_id_embedding.shape}')
         return uncond_id_embedding, id_embedding
 
-    def set_progress_bar_config(self, bar_format: str | None = None, ncols: int = 80, colour: str | None = None):
+    def set_progress_bar_config(self, bar_format: str | None = None, ncols: int = 120, colour: str | None = None):
         import functools
         from tqdm.auto import trange as trange_orig
         import pulid_sampling