From 5735b8f47d59ddf413a83ceb0ed10c8dccd56bfb Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 01:44:53 +0100
Subject: [PATCH 1/9] fix(ltx): pin guidance and cross-timestep on every 2.x
 call path

Some paths left both to pipeline defaults, which track the current upstream
model: distilled picked up guidance it already bakes in, and 2.0 could land on
2.3's joint-sigma path.
---
 modules/ltx/ltx_process.py | 40 +++++++++++++++++++++++++-------------
 1 file changed, 26 insertions(+), 14 deletions(-)

diff --git a/modules/ltx/ltx_process.py b/modules/ltx/ltx_process.py
index 39abe0643..5342bf597 100644
--- a/modules/ltx/ltx_process.py
+++ b/modules/ltx/ltx_process.py
@@ -32,11 +32,28 @@ def _prompt_tensors_to_device(*tensors):
     return tuple(t.to(device=devices.device) if torch.is_tensor(t) else t for t in tensors)
 
 
+def identity_ltx2_guidance() -> dict:
+    # Named rather than omitted: pipeline defaults track the current upstream model, so a missing
+    # term guides a schedule that already bakes it in.
+    return {
+        'stg_scale': 0.0,
+        'modality_scale': 1.0,
+        'guidance_rescale': 0.0,
+        'spatio_temporal_guidance_blocks': None,
+        'audio_guidance_scale': 1.0,
+        'audio_stg_scale': 0.0,
+        'audio_modality_scale': 1.0,
+        'audio_guidance_rescale': 0.0,
+    }
+
+
 def _canonical_ltx2_guidance(caps) -> dict:
     # Four-way composition (cfg + stg + modality + rescale) from huggingface/diffusers#13217.
-    # Distilled bakes these into its sigma schedule; skip or we double-apply.
-    if caps.family != '2.x' or caps.is_distilled:
+    # Distilled bakes these into its sigma schedule and runs at identity.
+    if caps.family != '2.x':
         return {}
+    if caps.is_distilled:
+        return identity_ltx2_guidance()
     return {
         'stg_scale': caps.stg_default_scale,
         'modality_scale': caps.modality_default_scale,
@@ -58,14 +75,7 @@ def _canonical_stage2_kwargs() -> dict:
         'sigmas': list(STAGE_2_DISTILLED_SIGMA_VALUES),
         'noise_scale': float(STAGE_2_DISTILLED_SIGMA_VALUES[0]),
         'guidance_scale': 1.0,
-        'stg_scale': 0.0,
-        'modality_scale': 1.0,
-        'guidance_rescale': 0.0,
-        'audio_guidance_scale': 1.0,
-        'audio_stg_scale': 0.0,
-        'audio_modality_scale': 1.0,
-        'audio_guidance_rescale': 0.0,
-        'spatio_temporal_guidance_blocks': None,
+        **identity_ltx2_guidance(),
     }
 
 
@@ -104,8 +114,8 @@ def _latent_pass(caps, prompt_embeds, prompt_attention_mask, negative_prompt_emb
         base_args['sigmas'] = list(DISTILLED_SIGMA_VALUES)
         base_args.pop('num_inference_steps', None)
     base_args.update(_canonical_ltx2_guidance(caps))
-    if caps.use_cross_timestep:
-        base_args['use_cross_timestep'] = True
+    if caps.family == '2.x':
+        base_args['use_cross_timestep'] = caps.use_cross_timestep
     log.debug(f'Video: cls={shared.sd_model.__class__.__name__} op=latent_pass args_keys={list(base_args.keys())}')
     result = shared.sd_model(**base_args)
     latents = result.frames[0] if hasattr(result, 'frames') else None
@@ -316,6 +326,8 @@ def run_ltx(task_id,
             p.task_args['sigmas'] = list(DISTILLED_SIGMA_VALUES)
             p.task_args.pop('num_inference_steps', None)
         p.task_args.update(_canonical_ltx2_guidance(caps))
+        if caps.family == '2.x':
+            p.task_args['use_cross_timestep'] = caps.use_cross_timestep
 
         framewise = caps.family == '0.9'
         set_vae_params(p, framewise=framewise)
@@ -499,8 +511,8 @@ def run_ltx(task_id,
                 # Thread Stage-1 I2V init image through Stage 2 so first-frame identity survives refine.
                 if caps.is_i2v and caps.repo_cls_name in ('LTXImageToVideoPipeline', 'LTX2ImageToVideoPipeline') and p.task_args.get('image') is not None:
                     refine_args['image'] = p.task_args['image']
-                if caps.family == '2.x' and caps.use_cross_timestep:
-                    refine_args['use_cross_timestep'] = True
+                if caps.family == '2.x':
+                    refine_args['use_cross_timestep'] = caps.use_cross_timestep
                 # output_type='latent' skips the post-loop audio_vae + vocoder pass when audio
                 # is unwanted; per-step audio cross-attention still runs for video conditioning.
                 # Internal video decode is also skipped; vae_decode below picks it up.

From 935c275d50cec4cb97050a70551a4dbcc62905ee Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 01:45:12 +0100
Subject: [PATCH 2/9] chore(deps): bump diffusers for ltx-2.5

Adds the Gemma 4 text encoder, the duration head and the diffusion decoder.
Pipeline classes are unchanged, so 2.0 and 2.3 load as before.

Image conditioning is now re-compressed to match training, at CRF 18 for 2.5
and 33 for earlier versions.
---
 installer.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/installer.py b/installer.py
index 4d1ce2534..eda04a677 100644
--- a/installer.py
+++ b/installer.py
@@ -584,7 +584,7 @@ def check_diffusers():
     t_start = time.time()
     if args.skip_all:
         return
-    target_commit = "90c0ffdc045902a3667d473d2fbfc03e8716dba9" # diffusers commit hash == 0.40.0.dev0 == 08-11-2026
+    target_commit = "7564fb016dabda0c943416190fc92398c50b1b20" # diffusers commit hash == 0.40.0.dev0 == 08-11-2026
     # if args.use_rocm or args.use_zluda:
     #     sha = '043ab2520f6a19fce78e6e060a68dbc947edb9f9' # lock diffusers versions for now
     pkg = package_spec('diffusers')

From 45fc1a0243db4371ff34652a180bc02483f61261 Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 01:53:19 +0100
Subject: [PATCH 3/9] fix(ltx): scope cached upsamplers and scheduler overrides
 to the run

The sampler shift override looked for flow_shift, which only UniPC-style
schedulers declare, so it never reached the flow-match schedulers 2.x uses.
Config keys are now restored only when the run wrote them, and a cached
upsampler is rebuilt once its repo or the model's VAE changes.
---
 modules/ltx/ltx_util.py | 83 +++++++++++++++++++++++++++--------------
 1 file changed, 56 insertions(+), 27 deletions(-)

diff --git a/modules/ltx/ltx_util.py b/modules/ltx/ltx_util.py
index 11afb0b1c..914c86834 100644
--- a/modules/ltx/ltx_util.py
+++ b/modules/ltx/ltx_util.py
@@ -37,7 +37,19 @@ def load_model(engine: str, model: str):
     timer.process.add('offload', t2 - t1)
 
 
+def upsample_pipe_stale(upsample_pipe, upsample_repo_id) -> bool:
+    # bound to one repo and borrowing the model's VAE; both change with the model, and a VAE from
+    # an unloaded model holds meta tensors
+    if upsample_pipe is None:
+        return False
+    if getattr(upsample_pipe, 'sdnext_upsample_repo', None) != upsample_repo_id:
+        return True
+    return upsample_pipe.vae is not getattr(shared.sd_model, 'vae', None)
+
+
 def load_upsample(upsample_pipe, upsample_repo_id):
+    if upsample_pipe_stale(upsample_pipe, upsample_repo_id):
+        upsample_pipe = None
     if upsample_pipe is None:
         t0 = time.time()
         from diffusers.pipelines.ltx.pipeline_ltx_latent_upsample import LTXLatentUpsamplePipeline
@@ -48,14 +60,17 @@ def load_upsample(upsample_pipe, upsample_repo_id):
             cache_dir=shared.opts.hfcache_dir,
             torch_dtype=devices.dtype,
         )
+        upsample_pipe.sdnext_upsample_repo = upsample_repo_id
         t1 = time.time()
         timer.process.add('load', t1 - t0)
     return upsample_pipe
 
 
-def load_upsample_2x(upsample_pipe, upsample_repo_id):
+def load_upsample_2x(upsample_pipe, upsample_repo_id, variant: str = '2.x'):
     # 2.x ships the upsampler as a bare nn.Module in a subfolder; no from_pretrained on the
     # pipeline wrapper, so we load the model + construct the pipeline manually.
+    if upsample_pipe_stale(upsample_pipe, upsample_repo_id):
+        upsample_pipe = None
     if upsample_pipe is None:
         t0 = time.time()
         from diffusers.pipelines.ltx2.pipeline_ltx2_latent_upsample import LTX2LatentUpsamplePipeline
@@ -74,40 +89,61 @@ def load_upsample_2x(upsample_pipe, upsample_repo_id):
         )
         # Synthetic checkpoint_info gives this pipe its own OffloadHook cache slot, so routing
         # it through apply_balanced_offload does not invalidate the main pipe's module map
-        # (sd_offload.py:488 keys on sd_checkpoint_info.name).
-        upsample_pipe.sd_checkpoint_info = sd_checkpoint.CheckpointInfo('ltx-upsampler-2.x')
+        # (sd_offload keys on sd_checkpoint_info.name). Variant is in the name since each loads
+        # different weights and the slot also names the disk offload folder.
+        upsample_pipe.sdnext_upsample_repo = upsample_repo_id
+        upsample_pipe.sd_checkpoint_info = sd_checkpoint.CheckpointInfo(f'ltx-upsampler-{variant}')
         t1 = time.time()
         timer.process.add('load', t1 - t0)
     return upsample_pipe
 
 
+def scheduler_shift_key(scheduler) -> str | None:
+    # UniPC and its relatives call the static shift flow_shift; flow-match schedulers call it shift.
+    config = getattr(scheduler, 'config', None)
+    if config is None:
+        return None
+    for key in ('flow_shift', 'shift'):
+        if hasattr(config, key):
+            return key
+    return None
+
+
 @contextmanager
-def ltx_scheduler_opts(sd_model, *, dynamic_shift=None, sampler_shift=None):
-    # Run-scoped override of shared.opts scheduler settings and scheduler.config. Snapshots
-    # five pieces of state (shared.opts dynamic_shift + shift, scheduler object, default_scheduler
-    # snapshot, and scheduler.config use_dynamic_shifting + flow_shift) and restores every one on
-    # exit. Keeps run-specific sampler settings out of config.json and prevents default_scheduler
-    # from getting clobbered by a deepcopy of the mutated scheduler at video_load.py:171. The
-    # scheduler-object restore matters for Stage 2 refine, which swaps the scheduler entirely.
+def ltx_scheduler_opts(sd_model, *, dynamic_shift=None, sampler_shift=None, shift_terminal=None):
+    # Run-scoped override of shared.opts scheduler settings and scheduler.config, restored on every
+    # exit path. Keeps run settings out of config.json, protects default_scheduler from a deepcopy
+    # of the mutated scheduler, and restores the scheduler object that Stage 2 refine swaps out.
     orig_dynamic_shift = shared.opts.schedulers_dynamic_shift
     orig_sampler_shift = shared.opts.schedulers_shift
     orig_scheduler = sd_model.scheduler
     orig_default_scheduler = getattr(sd_model, 'default_scheduler', None)
-    orig_use_dynamic_shifting = getattr(orig_scheduler.config, 'use_dynamic_shifting', None) if hasattr(orig_scheduler, 'config') else None
-    orig_flow_shift = getattr(orig_scheduler.config, 'flow_shift', None) if hasattr(orig_scheduler, 'config') else None
+    restore = {}
+
+    def write_config(values: dict):
+        scheduler = getattr(sd_model, 'scheduler', None)
+        if not values or scheduler is None or not hasattr(scheduler, 'config') or not hasattr(scheduler, 'register_to_config'):
+            return
+        for key, value in values.items():
+            setattr(scheduler.config, key, value)
+        scheduler.register_to_config(**values)
+
+    def override_config(key, value):
+        # only written keys are restored, so a key stored as None returns to None
+        if key is None or value is None or not hasattr(getattr(orig_scheduler, 'config', None), key):
+            return
+        restore[key] = getattr(orig_scheduler.config, key)
+        write_config({key: value})
 
     try:
         if dynamic_shift is not None:
             shared.opts.data['schedulers_dynamic_shift'] = dynamic_shift
         if sampler_shift is not None:
             shared.opts.data['schedulers_shift'] = sampler_shift
-        if hasattr(sd_model, 'scheduler') and hasattr(sd_model.scheduler, 'config') and hasattr(sd_model.scheduler, 'register_to_config'):
-            if dynamic_shift is not None and hasattr(sd_model.scheduler.config, 'use_dynamic_shifting'):
-                sd_model.scheduler.config.use_dynamic_shifting = dynamic_shift
-                sd_model.scheduler.register_to_config(use_dynamic_shifting=dynamic_shift)
-            if sampler_shift is not None and sampler_shift >= 0 and hasattr(sd_model.scheduler.config, 'flow_shift'):
-                sd_model.scheduler.config.flow_shift = sampler_shift
-                sd_model.scheduler.register_to_config(flow_shift=sampler_shift)
+        override_config('use_dynamic_shifting', dynamic_shift)
+        if sampler_shift is not None and sampler_shift >= 0:
+            override_config(scheduler_shift_key(orig_scheduler), sampler_shift)
+        override_config('shift_terminal', shift_terminal)
         yield
     finally:
         shared.opts.data['schedulers_dynamic_shift'] = orig_dynamic_shift
@@ -116,14 +152,7 @@ def ltx_scheduler_opts(sd_model, *, dynamic_shift=None, sampler_shift=None):
             sd_model.scheduler = orig_scheduler
         if orig_default_scheduler is not None and sd_model.default_scheduler is not orig_default_scheduler:
             sd_model.default_scheduler = orig_default_scheduler
-        if hasattr(sd_model.scheduler, 'config') and hasattr(sd_model.scheduler, 'register_to_config'):
-            if orig_use_dynamic_shifting is not None and hasattr(sd_model.scheduler.config, 'use_dynamic_shifting'):
-                sd_model.scheduler.config.use_dynamic_shifting = orig_use_dynamic_shifting
-                sd_model.scheduler.register_to_config(use_dynamic_shifting=orig_use_dynamic_shifting)
-            if orig_flow_shift is not None and hasattr(sd_model.scheduler.config, 'flow_shift'):
-                sd_model.scheduler.config.flow_shift = orig_flow_shift
-                sd_model.scheduler.register_to_config(flow_shift=orig_flow_shift)
-        # log.debug(f'LTX: scheduler/opts restored dynamic_shift={orig_dynamic_shift} sampler_shift={orig_sampler_shift}')
+        write_config(restore)
 
 
 def _condition_cls(family: str):

From 780495ff2e7686952e5f6a7b5a32bccde78c7404 Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 01:53:19 +0100
Subject: [PATCH 4/9] feat(video): allow a transformer folder to load into
 another argument

The subfolder name doubles as the pipeline argument, so a repo shipping a
second transformer beside the default one had no way to load it: the argument
was dropped and the default weights stayed. dit_kwarg names the argument when
it differs from the folder.
---
 modules/video_models/models_def.py |  1 +
 modules/video_models/video_load.py | 15 +++++++++------
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/modules/video_models/models_def.py b/modules/video_models/models_def.py
index 42b041a07..107dafab1 100644
--- a/modules/video_models/models_def.py
+++ b/modules/video_models/models_def.py
@@ -16,6 +16,7 @@ class Model:
     dit: str = None
     dit_cls: classmethod = None
     dit_folder: str = 'transformer'
+    dit_kwarg: str = None # pipeline argument the folder loads into, when the two differ
     dit_revision: str = None
     te: str = None
     te_cls: classmethod = None
diff --git a/modules/video_models/video_load.py b/modules/video_models/video_load.py
index 4e7415ae7..ff1a981b0 100644
--- a/modules/video_models/video_load.py
+++ b/modules/video_models/video_load.py
@@ -143,12 +143,13 @@ def load_model(selected: models_def.Model):
     # transformer
     if selected.dit_cls is not None:
         try:
-            def load_dit_folder(dit_folder):
-                if dit_folder is not None and dit_folder not in kwargs:
+            def load_dit_folder(dit_folder, dit_kwarg=None):
+                dit_kwarg = dit_kwarg or dit_folder # ltx-2.5 keeps its dev transformer in transformer_full
+                if dit_folder is not None and dit_kwarg not in kwargs:
                     # get a new quant arg on every loop to prevent the quant config classes getting entangled
                     load_args, quant_args = model_quant.get_dit_args({}, module='Model', device_map=True)
-                    log.debug(f'Load video: module=transformer repo="{selected.dit or selected.repo}" module="{dit_folder}" folder="{dit_folder}" cls={selected.dit_cls.__name__} quant={model_quant.get_quant_type(quant_args)} loader={_loader("diffusers")}')
-                    kwargs[dit_folder] = selected.dit_cls.from_pretrained(
+                    log.debug(f'Load video: module=transformer repo="{selected.dit or selected.repo}" module="{dit_kwarg}" folder="{dit_folder}" cls={selected.dit_cls.__name__} quant={model_quant.get_quant_type(quant_args)} loader={_loader("diffusers")}')
+                    kwargs[dit_kwarg] = selected.dit_cls.from_pretrained(
                         pretrained_model_name_or_path=selected.dit or selected.repo,
                         subfolder=dit_folder,
                         revision=selected.dit_revision or selected.repo_revision,
@@ -158,15 +159,17 @@ def load_model(selected: models_def.Model):
                         **offline_args,
                     )
                 else:
-                    log.debug(f'Load video: module=transformer repo="{selected.dit or selected.repo}" module="{dit_folder}" folder="{dit_folder}" cls={selected.dit_cls.__name__} loader={_loader("diffusers")} skip')
+                    log.debug(f'Load video: module=transformer repo="{selected.dit or selected.repo}" module="{dit_kwarg}" folder="{dit_folder}" cls={selected.dit_cls.__name__} loader={_loader("diffusers")} skip')
 
             if selected.dit_folder is None:
                 selected.dit_folder = ['transformer']
             if isinstance(selected.dit_folder, list) or isinstance(selected.dit_folder, tuple):
+                if selected.dit_kwarg is not None:
+                    log.warning(f'Load video: model="{selected.name}" dit_kwarg unsupported with multiple folders')
                 for dit_folder in selected.dit_folder: # wan a14b has transformer and transformer_2
                     load_dit_folder(dit_folder)
             else:
-                load_dit_folder(selected.dit_folder)
+                load_dit_folder(selected.dit_folder, selected.dit_kwarg)
         except Exception as e:
             log.error(f'video load: module=transformer cls={selected.dit_cls.__name__} {e}')
             errors.display(e, 'video')

From f01e752b062bb3343ea720b7433eb46811ff29cc Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 02:03:06 +0100
Subject: [PATCH 5/9] fix(video): take the audio rate from the loaded vocoder

Pipelines rarely report a sample rate, so the save path fell back to 24000.
LTX-2.3 and 2.5 run at 48k, and muxing at half the rate drops the track an
octave. The rate now comes from the vocoder, as the LTX tab already did.
---
 modules/video_models/video_run.py  |  2 +-
 modules/video_models/video_save.py | 10 ++++++++++
 2 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/modules/video_models/video_run.py b/modules/video_models/video_run.py
index 6444607eb..91e683e91 100644
--- a/modules/video_models/video_run.py
+++ b/modules/video_models/video_run.py
@@ -326,7 +326,7 @@ def run(selected: models_def.Model, *,
         p=p,
         pixels=pixels,
         audio=waveform,
-        aac_sample_rate=getattr(p, 'audio_sampling_rate', None) or 24000,
+        aac_sample_rate=video_save.get_audio_rate(p),
         binary=processed.bytes,
         mp4_fps=save_fps,
         mp4_codec=mp4_codec,
diff --git a/modules/video_models/video_save.py b/modules/video_models/video_save.py
index b14274ee6..323b67403 100644
--- a/modules/video_models/video_save.py
+++ b/modules/video_models/video_save.py
@@ -11,6 +11,16 @@ from modules.logger import log
 from modules.video_models.video_utils import check_av
 
 
+def get_audio_rate(p=None, default: int = 24000) -> int:
+    # pipeline output wins when it reports a rate, else the loaded vocoder: LTX-2.0 runs at 24k,
+    # 2.3 and 2.5 at 48k, and muxing at the wrong rate shifts the pitch
+    rate = getattr(p, 'audio_sampling_rate', None) if p is not None else None
+    if not rate:
+        vocoder = getattr(shared.sd_model, 'vocoder', None)
+        rate = getattr(getattr(vocoder, 'config', None), 'output_sampling_rate', None)
+    return int(rate) if rate else default
+
+
 def get_video_filename(p:processing.StableDiffusionProcessingVideo):
     from modules.image.namegen import FilenameGenerator
     from modules.paths import resolve_output_path

From 0d1882eca4ed047c882f0ae17e6953e6224ca88b Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 02:03:06 +0100
Subject: [PATCH 6/9] fix(video): keep text encoder dedup out of the registry
 rows

The shared repo was written back onto the registry row, a module-level
singleton, so turning the setting off left the row pointing at the shared copy
for the rest of the session. It is chosen into locals instead.
---
 modules/video_models/video_load.py | 52 ++++++++++++------------------
 1 file changed, 21 insertions(+), 31 deletions(-)

diff --git a/modules/video_models/video_load.py b/modules/video_models/video_load.py
index ff1a981b0..d1c94aca9 100644
--- a/modules/video_models/video_load.py
+++ b/modules/video_models/video_load.py
@@ -98,39 +98,29 @@ def load_model(selected: models_def.Model):
         try:
             load_args, quant_args = model_quant.get_dit_args({}, module='TE', device_map=True)
 
-            # loader deduplication of text-encoder models
-            if selected.te_cls.__name__ == 'T5EncoderModel' and shared.opts.te_shared_te:
-                selected.te = 'Disty0/t5-xxl'
-                selected.te_folder = ''
-                selected.te_revision = None
-            if selected.te_cls.__name__ == 'UMT5EncoderModel' and shared.opts.te_shared_te:
-                if 'SDNQ' in selected.name:
-                    selected.te = 'Disty0/Wan2.2-T2V-A14B-SDNQ-uint4-svd-r32'
-                else:
-                    selected.te = 'Wan-AI/Wan2.2-TI2V-5B-Diffusers'
-                selected.te_folder = 'text_encoder'
-                selected.te_revision = None
-            if selected.te_cls.__name__ == 'LlamaModel' and shared.opts.te_shared_te:
-                selected.te = 'hunyuanvideo-community/HunyuanVideo'
-                selected.te_folder = 'text_encoder'
-                selected.te_revision = None
-            if selected.te_cls.__name__ == 'Qwen2_5_VLForConditionalGeneration' and shared.opts.te_shared_te:
-                selected.te = 'ai-forever/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers'
-                selected.te_folder = 'text_encoder'
-                selected.te_revision = None
-            if selected.te_cls.__name__ == 'Gemma3ForConditionalGeneration' and shared.opts.te_shared_te:
-                if 'SDNQ' in selected.name:
-                    selected.te = 'OzzyGT/LTX-2.3-sdnq-dynamic-int4'
-                else:
-                    selected.te = 'OzzyGT/LTX-2.3'
-                selected.te_folder = 'text_encoder'
-                selected.te_revision = None
+            # loader deduplication of text-encoder models: picked per load, not written back onto
+            # the registry row where it would outlive the setting
+            te_repo, te_folder, te_revision = selected.te, selected.te_folder, selected.te_revision
+            if shared.opts.te_shared_te:
+                te_cls_name = selected.te_cls.__name__
+                if te_cls_name == 'T5EncoderModel':
+                    te_repo, te_folder, te_revision = 'Disty0/t5-xxl', '', None
+                elif te_cls_name == 'UMT5EncoderModel':
+                    te_repo = 'Disty0/Wan2.2-T2V-A14B-SDNQ-uint4-svd-r32' if 'SDNQ' in selected.name else 'Wan-AI/Wan2.2-TI2V-5B-Diffusers'
+                    te_folder, te_revision = 'text_encoder', None
+                elif te_cls_name == 'LlamaModel':
+                    te_repo, te_folder, te_revision = 'hunyuanvideo-community/HunyuanVideo', 'text_encoder', None
+                elif te_cls_name == 'Qwen2_5_VLForConditionalGeneration':
+                    te_repo, te_folder, te_revision = 'ai-forever/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers', 'text_encoder', None
+                elif te_cls_name == 'Gemma3ForConditionalGeneration':
+                    te_repo = 'OzzyGT/LTX-2.3-sdnq-dynamic-int4' if 'SDNQ' in selected.name else 'OzzyGT/LTX-2.3'
+                    te_folder, te_revision = 'text_encoder', None
 
-            log.debug(f'Load video: module=te repo="{selected.te or selected.repo}" folder="{selected.te_folder}" cls={selected.te_cls.__name__} quant={model_quant.get_quant_type(quant_args)} loader={_loader("transformers")}')
+            log.debug(f'Load video: module=te repo="{te_repo or selected.repo}" folder="{te_folder}" cls={selected.te_cls.__name__} quant={model_quant.get_quant_type(quant_args)} loader={_loader("transformers")}')
             kwargs["text_encoder"] = selected.te_cls.from_pretrained(
-                pretrained_model_name_or_path=selected.te or selected.repo,
-                subfolder=selected.te_folder,
-                revision=selected.te_revision or selected.repo_revision,
+                pretrained_model_name_or_path=te_repo or selected.repo,
+                subfolder=te_folder,
+                revision=te_revision or selected.repo_revision,
                 cache_dir=shared.opts.hfcache_dir,
                 **load_args,
                 **quant_args,

From 5814d5c4b3cd2db8ca31aec5983a7729de121dbb Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 02:03:18 +0100
Subject: [PATCH 7/9] feat(ltx): add the ltx-2.5 model family

2.5 reuses the LTX-2 pipeline classes, so it is described through the
capability table: Gemma 4 text encoder, cross-timestep conditioning, and the
upsampler and stage 2 LoRA that now ship inside the model repo. The repo
carries the distilled scheduler config, so Dev rows restore the terminal
shift, and the Dev transformer sits in transformer_full.

Distilled rows stop forcing dynamic shifting on, which remapped their sigma
schedule. Auto duration hands the clip length to the duration head.
---
 CHANGELOG.md                            | 14 ++++++++
 modules/ltx/ltx_capabilities.py         | 40 ++++++++++++++++-----
 modules/ltx/ltx_process.py              | 40 ++++++++++++---------
 modules/ltx/ltx_ui.py                   |  8 ++++-
 modules/video_models/models_def.py      | 46 +++++++++++++++++++++++++
 modules/video_models/video_overrides.py | 21 +++++++----
 ui/locale/locale_en.json                |  9 ++---
 7 files changed, 140 insertions(+), 38 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 681cc1331..47e1f166b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,14 @@
   - [MiniMax H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) in *base* and *ref* variants
     MiniMax-H3 is an amazing, but absolutely massive at 32B text-encoder and 33B transformer video model  
     for details, see [MiniMax wiki page](wiki/MiniMax)
+  - [LTX-2.5](https://huggingface.co/Lightricks/LTX-2.5) from Lightricks  
+    22B joint audio and video generation in *Distilled* and *Dev* variants,  
+    each as *text-to-video*, *image-to-video* and *conditioning* rows  
+    paired with a *Gemma 4* text encoder and a duration head:  
+    tick *auto duration* and the clip length is predicted from the prompt instead of set by hand  
+    *note*: LTX-2.5 is a [gated model](https://vladmandic.github.io/sdnext-docs/Gated/)  
+    *note*: image conditioning is now re-compressed to match what the models were trained on,  
+    at CRF 18 for LTX-2.5 and 33 for earlier LTX-2.x versions  
 - **Detailer**: Pretty much *detailer.next* :)  
   Detailer detection models were traditionally *YOLO* models, but now we can also use:  
   - [Facebook-SAM3](https://huggingface.co/facebook/sam3) hybrid promptable concept segmentation and detection network
@@ -43,6 +51,12 @@
   - improve pipeline detection for non-cached models
   - cleanup alt offload codepaths
   - hf progress bars
+  - ltx: send the guidance stack and cross-timestep on every 2.x call path
+  - ltx: distilled variants no longer force dynamic shifting on, which remapped their sigma schedule
+  - ltx: sampler shift now reaches flow-match schedulers
+  - ltx: reload the latent upsampler when the model or its repo changes
+  - video: take the audio sample rate from the loaded vocoder
+  - video: keep the shared text encoder out of the registry rows
 
 ## Update for 2026-08-07
 
diff --git a/modules/ltx/ltx_capabilities.py b/modules/ltx/ltx_capabilities.py
index 0d1909580..662d8c6d9 100644
--- a/modules/ltx/ltx_capabilities.py
+++ b/modules/ltx/ltx_capabilities.py
@@ -9,7 +9,7 @@ class LTXCaps:
     name: str
     repo_cls_name: str
     family: str  # '0.9' or '2.x'
-    variant: str  # '0.9', '2.0', '2.3' (finer-grained sub-variant)
+    variant: str  # '0.9', '2.0', '2.3', '2.5' (finer-grained sub-variant)
     is_distilled: bool
     is_i2v: bool
     supports_input_media: bool
@@ -41,6 +41,13 @@ class LTXCaps:
     # for refine; same-res refine oversaturates. Condition variants rebuild conditions per stage.
     supports_two_stage_refine: bool = False
     stage2_dev_lora_repo: Optional[str] = None
+    # 2.5 keeps the stage 2 LoRA in the model repo; 2.0 and 2.3 each have their own
+    stage2_dev_lora_weight: Optional[str] = None
+    # tied to the family VAE: the wrong one drifts per-channel latent statistics
+    upsample_repo: Optional[str] = None
+    # 2.5 ships the distilled scheduler config, so its Dev rows restore the terminal shift
+    scheduler_shift_terminal: Optional[float] = None
+    supports_auto_duration: bool = False
 
 
 CONDITION_CLASSES = {'LTXConditionPipeline', 'LTX2ConditionPipeline'}
@@ -84,7 +91,12 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
     family = '2.x' if is_ltx2 else '0.9'
     # 2.x sub-variant detection: unknown 2.x mirrors fall through to '2.0' (conservative default).
     if is_ltx2:
-        variant = '2.3' if '2.3' in model_name else '2.0'
+        if '2.5' in model_name:
+            variant = '2.5'
+        elif '2.3' in model_name:
+            variant = '2.3'
+        else:
+            variant = '2.0'
     else:
         variant = '0.9'
     is_distilled = 'Distilled' in model_name
@@ -106,11 +118,12 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
         supports_stg=is_ltx2,
         supports_audio=is_ltx2,
         supports_frame_rate_kwarg=is_ltx2,
-        use_cross_timestep=(variant == '2.3'),
+        use_cross_timestep=variant in ('2.3', '2.5'),
         default_cfg=3.0,
         default_steps=30 if is_ltx2 else 50,
         default_sampler_shift=-1.0,
-        default_dynamic_shift=is_ltx2,
+        # distilled ships use_dynamic_shifting=False and runs explicit sigmas, which shifting remaps
+        default_dynamic_shift=is_ltx2 and not is_distilled,
         default_width=768,
         default_height=512,
         default_frames=121 if is_ltx2 else 161,
@@ -122,7 +135,10 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
         caps.default_steps = 8
 
     if is_ltx2 and not is_distilled:
-        if variant == '2.3':
+        if variant == '2.5':
+            caps.stage2_dev_lora_repo = 'Lightricks/LTX-2.5-Diffusers'
+            caps.stage2_dev_lora_weight = 'ltx-2.5-22b-distilled-lora-450-bf16.safetensors'
+        elif variant == '2.3':
             caps.stage2_dev_lora_repo = 'CalamitousFelicitousness/LTX-2.3-distilled-lora-384-Diffusers'
         elif variant == '2.0':
             caps.stage2_dev_lora_repo = 'CalamitousFelicitousness/LTX-2.0-distilled-lora-384-Diffusers'
@@ -130,16 +146,22 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
     caps.supports_two_stage_refine = is_ltx2
 
     if is_ltx2:
-        if variant == '2.3':
+        if variant == '2.5':
+            caps.upsample_repo = 'Lightricks/LTX-2.5-Diffusers'
+            caps.stg_default_blocks = [28]
+            caps.supports_auto_duration = True
+        elif variant == '2.3':
+            caps.upsample_repo = 'CalamitousFelicitousness/LTX-2.3-Spatial-Upsampler-x2-1.1-Diffusers'
             caps.stg_default_blocks = [28]
-        elif variant == '2.0':
-            caps.stg_default_blocks = [29]
         else:
-            caps.stg_default_blocks = [28]
+            caps.upsample_repo = 'Lightricks/LTX-2'
+            caps.stg_default_blocks = [29]
         if not is_distilled:
             # canonical T2V composition from huggingface/diffusers#13217
             caps.stg_default_scale = 1.0
             caps.modality_default_scale = 3.0
             caps.guidance_rescale_default = 0.7
+            if variant == '2.5':
+                caps.scheduler_shift_terminal = 0.1
 
     return caps
diff --git a/modules/ltx/ltx_process.py b/modules/ltx/ltx_process.py
index 5342bf597..c684dcee2 100644
--- a/modules/ltx/ltx_process.py
+++ b/modules/ltx/ltx_process.py
@@ -12,16 +12,12 @@ from modules.ltx.ltx_util import get_bucket, get_frames, load_model, load_upsamp
 apply_ltx_diffusers_patch()
 from modules.processing_callbacks import diffusers_callback
 from modules.video_models.video_vae import set_vae_params
-from modules.video_models.video_save import save_video
+from modules.video_models.video_save import save_video, get_audio_rate
 from modules.video_models.video_utils import check_av
 
 
 debug = log.trace if os.environ.get('SD_VIDEO_DEBUG', None) is not None else lambda *args, **kwargs: None
 upsample_repo_id_09 = 'a-r-r-o-w/LTX-Video-0.9.7-Latent-Spatial-Upsampler-diffusers'
-# Upsampler weights are tied to the family VAE; using the wrong one preserves structure
-# but drifts per-channel latent statistics (decodes desaturated / crushed contrast).
-upsample_repo_id_20 = 'Lightricks/LTX-2'
-upsample_repo_id_23 = 'CalamitousFelicitousness/LTX-2.3-Spatial-Upsampler-x2-1.1-Diffusers'
 upsample_pipe = None
 upsample_pipe_2x = None
 
@@ -93,7 +89,7 @@ def _latent_pass(caps, prompt_embeds, prompt_attention_mask, negative_prompt_emb
         'negative_prompt_attention_mask': negative_prompt_attention_mask,
         'width': get_bucket(width),
         'height': get_bucket(height),
-        'num_frames': get_frames(frames),
+        'num_frames': get_frames(frames) if frames is not None else None, # None defers to the duration head
         'num_inference_steps': steps,
         'generator': get_generator(seed),
         'callback_on_step_end': diffusers_callback,
@@ -131,6 +127,7 @@ def run_ltx(task_id,
             width: int,
             height: int,
             frames: int,
+            auto_duration: bool,
             steps: int,
             sampler_index: int,
             guidance_scale: float,
@@ -199,6 +196,10 @@ def run_ltx(task_id,
             yield from abort(f'Video: cls={shared.sd_model.__class__.__name__} selected model is not LTX', ok=True)
             return
 
+        auto_frames = bool(auto_duration) and caps.supports_auto_duration
+        if auto_duration and not auto_frames:
+            log.warning(f'LTX: model="{model}" auto duration unsupported, using frames={get_frames(frames)}')
+
         # Lightricks TI2VidTwoStagesPipeline: Stage 1 at half-res, 2x upsample, Stage 2 refine at target.
         # Auto-couple when the user picks Refine but not Upsample. Both Dev and Distilled refine paths
         # expect upsampled latents; same-res refine on Distilled produces oversaturation. Condition
@@ -328,14 +329,15 @@ def run_ltx(task_id,
         p.task_args.update(_canonical_ltx2_guidance(caps))
         if caps.family == '2.x':
             p.task_args['use_cross_timestep'] = caps.use_cross_timestep
+        if auto_frames:
+            p.task_args['num_frames'] = None
 
         framewise = caps.family == '0.9'
         set_vae_params(p, framewise=framewise)
 
         # Scheduler + shared.opts mutation is wrapped in ltx_scheduler_opts so restore runs on
-        # every exit path (normal return, abort, interrupt, Stage 2 scheduler swap). See the
-        # helper's docstring for the five pieces of state it snapshots.
-        with ltx_scheduler_opts(shared.sd_model, dynamic_shift=dynamic_shift, sampler_shift=sampler_shift):
+        # every exit path (normal return, abort, interrupt, Stage 2 scheduler swap).
+        with ltx_scheduler_opts(shared.sd_model, dynamic_shift=dynamic_shift, sampler_shift=sampler_shift, shift_terminal=caps.scheduler_shift_terminal):
             if selected is not None:
                 video_overrides.set_overrides(p, selected)
 
@@ -382,7 +384,7 @@ def run_ltx(task_id,
                         negative_prompt_attention_mask=negative_prompt_attention_mask,
                         width=base_w,
                         height=base_h,
-                        frames=frames,
+                        frames=None if auto_frames else frames,
                         steps=steps,
                         guidance_scale=p.cfg_scale,
                         mp4_fps=mp4_fps,
@@ -391,6 +393,11 @@ def run_ltx(task_id,
                         seed=p.seed,
                         image=p.task_args.get('image'),
                     )
+                    if auto_frames and torch.is_tensor(latents):
+                        # upsample and refine take the realized length; re-predicting would drift
+                        frames = (latents.shape[-3] - 1) * getattr(shared.sd_model, 'vae_temporal_compression_ratio', 8) + 1
+                        p.frames = frames
+                        log.debug(f'LTX: auto duration frames={frames}')
                 else:
                     processed = processing.process_images(p)
                     if processed is None or processed.images is None or len(processed.images) == 0:
@@ -451,8 +458,7 @@ def run_ltx(task_id,
                         upsample_pipe = sd_models.apply_balanced_offload(upsample_pipe, exclude=upsample_exclude, silent=True)
                     else:
                         global upsample_pipe_2x # pylint: disable=global-statement
-                        upsample_repo = upsample_repo_id_23 if caps.variant == '2.3' else upsample_repo_id_20
-                        upsample_pipe_2x = load_upsample_2x(upsample_pipe_2x, upsample_repo)
+                        upsample_pipe_2x = load_upsample_2x(upsample_pipe_2x, caps.upsample_repo, caps.variant)
                         upsample_pipe_2x = sd_models.apply_balanced_offload(upsample_pipe_2x, exclude=upsample_exclude, silent=True)
                         # 2.x base pass returns denormalized latents; latents_normalized=False tells the
                         # upsampler "already raw, do not denormalize again".
@@ -534,12 +540,15 @@ def run_ltx(task_id,
                             shift_terminal=None,
                         )
                         if caps.supports_canonical_stage2:
-                            log.debug(f'LTX: stage=2 distilled=LoRA repo={caps.stage2_dev_lora_repo}')
+                            log.debug(f'LTX: stage=2 distilled=LoRA repo={caps.stage2_dev_lora_repo} weight={caps.stage2_dev_lora_weight}')
                             offline_args = {'local_files_only': True} if shared.opts.offline_mode else {}
+                            # 2.5 keeps the LoRA in the model repo, so the file has to be named
+                            lora_args ={'weight_name': caps.stage2_dev_lora_weight} if caps.stage2_dev_lora_weight is not None else {}
                             shared.sd_model.load_lora_weights(
                                 caps.stage2_dev_lora_repo,
                                 adapter_name=STAGE2_DEV_LORA_ADAPTER,
                                 cache_dir=shared.opts.hfcache_dir,
+                                **lora_args,
                                 **offline_args,
                             )
                             shared.sd_model.set_adapters([STAGE2_DEV_LORA_ADAPTER], [1.0])
@@ -628,10 +637,7 @@ def run_ltx(task_id,
             if not audio_enable:
                 audio = None
 
-            try:
-                aac_sample_rate = shared.sd_model.vocoder.config.output_sampling_rate
-            except Exception:
-                aac_sample_rate = 24000
+            aac_sample_rate = get_audio_rate(p)
 
             if mp4_interpolate > 0 and pixels is not None:
                 p.video_interpolate = mp4_interpolate
diff --git a/modules/ltx/ltx_ui.py b/modules/ltx/ltx_ui.py
index 3d381933f..550b79686 100644
--- a/modules/ltx/ltx_ui.py
+++ b/modules/ltx/ltx_ui.py
@@ -29,6 +29,7 @@ def _model_change(model_name: str):
             gr.update(interactive=False),  # decode_timestep
             gr.update(interactive=False),  # image_cond_noise_scale
             gr.update(visible=False),   # audio_accordion
+            gr.update(visible=False, value=False),  # auto_duration
         )
     # 2.x refine runs fixed canonical schedules; refine_strength only feeds 0.9.x LTXConditionPipeline.
     refine_strength_interactive = caps.family == '0.9'
@@ -36,6 +37,7 @@ def _model_change(model_name: str):
     # Distilled T2V/I2V). auto_refine_upsample at ltx_process.py:179 couples the stages once Refine
     # is on. Condition variants are excluded by supports_two_stage_refine.
     refine_default = caps.supports_two_stage_refine
+    auto_duration_update =gr.update(visible=True) if caps.supports_auto_duration else gr.update(visible=False, value=False)
     return (
         gr.update(visible=caps.supports_input_media),
         gr.update(visible=caps.supports_multi_condition),
@@ -52,6 +54,7 @@ def _model_change(model_name: str):
         gr.update(interactive=caps.supports_decode_timestep),
         gr.update(interactive=caps.supports_image_cond_noise_scale),
         gr.update(visible=caps.supports_audio),
+        auto_duration_update,
     )
 
 
@@ -71,6 +74,8 @@ def create_ui(prompt, negative, styles, overrides, script_inputs, mp4_fps, mp4_i
                     seed = gr.Number(label='LTX seed', value=-1, elem_id='ltx_seed', container=True)
                     random_seed = ToolButton(ui_symbols.random, elem_id='ltx_seed_random')
                     random_seed.click(fn=lambda: -1, show_progress='hidden', inputs=[], outputs=[seed])
+                with gr.Row():
+                    auto_duration = gr.Checkbox(label='LTX auto duration', value=False, elem_id='ltx_auto_duration', visible=False)
             input_media_accordion = gr.Accordion(open=False, label="Input media", elem_id='ltx_input_media_accordion', visible=False)
             with input_media_accordion:
                 ltx_init_image = gr.Image(label='Image', elem_id='ltx_init_image', type='pil', image_mode='RGB', width=256, height=256)
@@ -145,6 +150,7 @@ def create_ui(prompt, negative, styles, overrides, script_inputs, mp4_fps, mp4_i
             decode_timestep,
             image_cond_noise_scale,
             audio_accordion,
+            auto_duration,
         ],
     )
 
@@ -155,7 +161,7 @@ def create_ui(prompt, negative, styles, overrides, script_inputs, mp4_fps, mp4_i
     video_inputs = [
         model,
         prompt, negative, styles,
-        width, height, frames,
+        width, height, frames, auto_duration,
         steps, sampler_index,
         guidance_scale, sampler_shift, dynamic_shift,
         seed,
diff --git a/modules/video_models/models_def.py b/modules/video_models/models_def.py
index 107dafab1..12dcdc27d 100644
--- a/modules/video_models/models_def.py
+++ b/modules/video_models/models_def.py
@@ -149,6 +149,52 @@ try:
         'LTX Video': [
             Model(name='None'),
 
+            Model(name='─────── LTX-2.5 ───────'),
+            Model(name='─── Distilled ───'),
+            Model(name='LTXVideo 2.5 22B T2V Distilled',
+                url='https://huggingface.co/Lightricks/LTX-2.5',
+                repo='Lightricks/LTX-2.5-Diffusers',
+                repo_cls='LTX2Pipeline',
+                te_cls='Gemma4UnifiedForConditionalGeneration',
+                dit_cls='LTX2VideoTransformer3DModel'),
+            Model(name='LTXVideo 2.5 22B I2V Distilled',
+                url='https://huggingface.co/Lightricks/LTX-2.5',
+                repo='Lightricks/LTX-2.5-Diffusers',
+                repo_cls='LTX2ImageToVideoPipeline',
+                te_cls='Gemma4UnifiedForConditionalGeneration',
+                dit_cls='LTX2VideoTransformer3DModel'),
+            Model(name='LTXVideo 2.5 22B Condition Distilled',
+                url='https://huggingface.co/Lightricks/LTX-2.5',
+                repo='Lightricks/LTX-2.5-Diffusers',
+                repo_cls='LTX2ConditionPipeline',
+                te_cls='Gemma4UnifiedForConditionalGeneration',
+                dit_cls='LTX2VideoTransformer3DModel'),
+            Model(name='─── Dev ───'), # transformer_full is the guided model, transformer the distilled one
+            Model(name='LTXVideo 2.5 22B T2V Dev',
+                url='https://huggingface.co/Lightricks/LTX-2.5',
+                repo='Lightricks/LTX-2.5-Diffusers',
+                repo_cls='LTX2Pipeline',
+                te_cls='Gemma4UnifiedForConditionalGeneration',
+                dit_cls='LTX2VideoTransformer3DModel',
+                dit_folder='transformer_full',
+                dit_kwarg='transformer'),
+            Model(name='LTXVideo 2.5 22B I2V Dev',
+                url='https://huggingface.co/Lightricks/LTX-2.5',
+                repo='Lightricks/LTX-2.5-Diffusers',
+                repo_cls='LTX2ImageToVideoPipeline',
+                te_cls='Gemma4UnifiedForConditionalGeneration',
+                dit_cls='LTX2VideoTransformer3DModel',
+                dit_folder='transformer_full',
+                dit_kwarg='transformer'),
+            Model(name='LTXVideo 2.5 22B Condition Dev',
+                url='https://huggingface.co/Lightricks/LTX-2.5',
+                repo='Lightricks/LTX-2.5-Diffusers',
+                repo_cls='LTX2ConditionPipeline',
+                te_cls='Gemma4UnifiedForConditionalGeneration',
+                dit_cls='LTX2VideoTransformer3DModel',
+                dit_folder='transformer_full',
+                dit_kwarg='transformer'),
+
             Model(name='─────── LTX-2.3 v1.1 ───────'),
             Model(name='LTXVideo 2.3-1.1 22B T2V Distilled',
                 url='https://huggingface.co/Lightricks/LTX-2.3',
diff --git a/modules/video_models/video_overrides.py b/modules/video_models/video_overrides.py
index 69afe7fb4..1964bdc59 100644
--- a/modules/video_models/video_overrides.py
+++ b/modules/video_models/video_overrides.py
@@ -17,17 +17,24 @@ def load_override(selected: Model, **load_args):
     # LTX
     if 'LTXVideo 0.9.5 I2V' in selected.name:
         kwargs['vae'] = diffusers.AutoencoderKLLTXVideo.from_pretrained(selected.repo, subfolder="vae", torch_dtype=torch.float32, cache_dir=shared.opts.hfcache_dir, **load_args)
-    # OzzyGT LTX-2.3 mirrors pack connectors/ twice by design: sharded (*-00001-of-0000N +
-    # .index.json) and unsharded diffusion_pytorch_model.safetensors of the byte-identical
-    # weights. snapshot_download faithfully fetches both; diffusers' component loader picks
-    # sharded when the index is present. ignore_patterns skips the ~6.3 GB unsharded copy
-    # without reaching for a cleaner upstream mirror.
+    # OzzyGT LTX-2.3 mirrors and the LTX-2.5 repo pack connectors/ twice by design: sharded
+    # (*-00001-of-0000N + .index.json) and unsharded diffusion_pytorch_model.safetensors of the
+    # byte-identical weights. snapshot_download faithfully fetches both; diffusers' component
+    # loader picks sharded when the index is present. ignore_patterns skips the ~6.3 GB unsharded
+    # copy without reaching for a cleaner upstream mirror.
     ltx2_redundant_connector_repos = {
         'OzzyGT/LTX-2.3',
         'OzzyGT/LTX-2.3-sdnq-dynamic-int4',
     }
-    if selected.repo in ltx2_redundant_connector_repos:
-        kwargs['ignore_patterns'] = ['connectors/diffusion_pytorch_model.safetensors']
+    ltx2_ignore = []
+    if selected.repo in ltx2_redundant_connector_repos or 'LTXVideo 2.5' in selected.name:
+        ltx2_ignore.append('connectors/diffusion_pytorch_model.safetensors')
+    if 'LTXVideo 2.5' in selected.name:
+        # the pipeline fetch pulls every model-index folder except passed components: transformer_full
+        # is not one, the diffusion decoder is a separate pipeline, the LoRA is fetched on demand
+        ltx2_ignore += ['transformer_full/*', 'diffusion_decoder/*', 'ltx-2.5-22b-distilled-lora-450-bf16.safetensors']
+    if ltx2_ignore:
+        kwargs['ignore_patterns'] = ltx2_ignore
     # LTX2TextConnectors weights are byte-identical across all 2.3 variants (verified by blob
     # hash). Pre-load from a canonical repo so per-variant fetches skip connectors/ entirely.
     # FP16 variants share OzzyGT/LTX-2.3; SDNQ variants share the pre-quantized mirror.
diff --git a/ui/locale/locale_en.json b/ui/locale/locale_en.json
index 532f1c81c..b216e8c41 100644
--- a/ui/locale/locale_en.json
+++ b/ui/locale/locale_en.json
@@ -832,10 +832,11 @@
     {"id":"","label":"LTX model","localized":"","hint":"","ui":"video"},
     {"id":"","label":"LTX frames number","localized":"","hint":"","ui":"video"},
     {"id":"","label":"LTX frames skip","localized":"","hint":"","ui":"video"},
-    {"id":"","label":"LTX enable upsampling","localized":"","hint":"","ui":"video"},
-    {"id":"","label":"LTX upsample ratio","localized":"","hint":"","ui":"video"},
-    {"id":"","label":"LTX enable refine","localized":"","hint":"","ui":"video"},
-    {"id":"","label":"LTX refine strength","localized":"","hint":"","ui":"video"},
+    {"id":"","label":"LTX auto duration","localized":"","hint":"Clip length is predicted from the prompt and the frames setting is ignored","ui":"video"},
+    {"id":"","label":"LTX upscale","localized":"","hint":"","ui":"video"},
+    {"id":"","label":"LTX scale","localized":"","hint":"","ui":"video"},
+    {"id":"","label":"LTX refine","localized":"","hint":"","ui":"video"},
+    {"id":"","label":"LTX strength","localized":"","hint":"","ui":"video"},
     {"id":"","label":"LTX decode timestep","localized":"","hint":"","ui":"video"},
     {"id":"","label":"LTX save audio","localized":"","hint":"LTX-2 audio-capable models always generate audio from the same prompt as video; this toggle controls whether the audio track is included in the saved video file","ui":"video"},
     {"id":"","label":"Loop","localized":"","hint":"","ui":"extras"},

From 5cb6efea340e9c63a4f1f2d5f18c712e3e4e4548 Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Thu, 13 Aug 2026 02:36:40 +0100
Subject: [PATCH 8/9] fix(ltx): apply the 2.5 stage-2 lora to the text
 connectors

The stage 2 LoRA stores its connector deltas as
diffusion_model.{video,audio}_embeddings_connector, but
LTX2LoraLoaderMixin.lora_state_dict recognizes connectors only under the 2.3
text_embedding_projection prefix. All 3544 keys are routed into the transformer
namespace and peft drops the 224 that land nowhere, leaving refine with a
transformer-only adapter.

Wrapping lora_state_dict moves those keys onto the connectors component using
the rename table from the convert_ltx2_to_diffusers script. The wrapper is inert
once no misrouted keys appear, so it needs no version check.
---
 modules/ltx/ltx_diffusers_patch.py | 103 ++++++++++++++++++++++-------
 1 file changed, 80 insertions(+), 23 deletions(-)

diff --git a/modules/ltx/ltx_diffusers_patch.py b/modules/ltx/ltx_diffusers_patch.py
index 4444cb19f..43bef4ce3 100644
--- a/modules/ltx/ltx_diffusers_patch.py
+++ b/modules/ltx/ltx_diffusers_patch.py
@@ -1,29 +1,47 @@
-"""Workaround for huggingface/diffusers#13564 connectors padding regression.
+"""Local fixes for LTX-2.x gaps in the pinned diffusers.
 
-PR #13564 (merged 2026-05-08) refactored LTX2ConnectorTransformer1d's padding
-logic from a loop-based gather-and-pad into a vectorized mask-then-flip. The
-new code applies torch.flip(hidden_states, dims=[1]) after replacing padding
-positions with learned registers, which reverses the order of valid prompt
-tokens. Audio cross-attention is position-sensitive, so reversed token order
-produces jumbled dialogue (right vocabulary, wrong word order). Visual quality
-is mostly unaffected because spatial cross-attention is less position-sensitive.
+Both patches are installed at import time by ltx_process and are safe to leave in
+place once upstream fixes them: the first skips when the source no longer matches,
+the second is a no-op as soon as no misrouted keys appear.
 
-This module restores the pre-#13564 forward at import time when the broken
-pattern is detected. Safe to leave in place after upstream fixes the bug:
-detection will skip the monkey-patch when the source no longer matches.
+Connector padding (huggingface/diffusers#13564): PR #13564 (merged 2026-05-08)
+refactored LTX2ConnectorTransformer1d's padding logic from a loop-based
+gather-and-pad into a vectorized mask-then-flip. The new code applies
+torch.flip(hidden_states, dims=[1]) after replacing padding positions with learned
+registers, which reverses the order of valid prompt tokens. Audio cross-attention is
+position-sensitive, so reversed token order produces jumbled dialogue (right
+vocabulary, wrong word order). Visual quality is mostly unaffected because spatial
+cross-attention is less position-sensitive.
+
+Stage-2 LoRA connectors: LTX2LoraLoaderMixin.lora_state_dict recognizes connector
+weights only under the 2.3-era text_embedding_projection prefix, so a
+diffusion_model.* checkpoint is routed wholesale into the transformer namespace. The
+2.5 stage-2 distilled LoRA carries its connector deltas as
+diffusion_model.{video,audio}_embeddings_connector.*, so 224 of its 3544 keys reach a
+module that cannot host them and peft drops them. Re-routing uses the rename table
+from the convert_ltx2_to_diffusers script.
 """
 
+import functools
 import inspect
 
 import torch
 import torch.nn.functional as F
 
-
-_PATCH_APPLIED = False
-_BROKEN_MARKER = 'torch.flip(hidden_states, dims=[1])'
+from modules.logger import log
 
 
-def _patched_forward(
+PATCH_APPLIED = False
+BROKEN_MARKER = 'torch.flip(hidden_states, dims=[1])'
+CONNECTOR_LORA_PREFIXES = ('video_embeddings_connector.', 'audio_embeddings_connector.')
+CONNECTOR_LORA_RENAME = {
+    'video_embeddings_connector': 'video_connector',
+    'audio_embeddings_connector': 'audio_connector',
+    'transformer_1d_blocks': 'transformer_blocks',
+}
+
+
+def patched_connector_forward(
     self,
     hidden_states: torch.Tensor,
     attention_mask: torch.Tensor | None = None,
@@ -72,19 +90,58 @@ def _patched_forward(
     return hidden_states, attention_mask
 
 
-def apply_patch():
-    global _PATCH_APPLIED # pylint: disable=global-statement
-    if _PATCH_APPLIED:
-        return
+def reroute_connector_keys(state_dict):
+    converted = {}
+    moved = 0
+    for key, value in state_dict.items():
+        name = key.removeprefix('transformer.')
+        if name.startswith(CONNECTOR_LORA_PREFIXES):
+            for src, dst in CONNECTOR_LORA_RENAME.items():
+                name = name.replace(src, dst)
+            converted[f'connectors.{name}'] = value
+            moved += 1
+        else:
+            converted[key] = value
+    if moved == 0:
+        return state_dict
+    log.debug(f'LTX: lora=connectors rerouted={moved} total={len(state_dict)}')
+    return converted
+
+
+def apply_connectors_forward_patch():
     try:
         from diffusers.pipelines.ltx2.connectors import LTX2ConnectorTransformer1d
     except ImportError:
-        _PATCH_APPLIED = True
         return
     try:
         source = inspect.getsource(LTX2ConnectorTransformer1d.forward)
     except (OSError, TypeError):
         source = ''
-    if _BROKEN_MARKER in source:
-        LTX2ConnectorTransformer1d.forward = _patched_forward # TODO ltx: patched diffusers connectors padding to fix audio token order (upstream #13564 regression)
-    _PATCH_APPLIED = True
+    if BROKEN_MARKER in source:
+        LTX2ConnectorTransformer1d.forward = patched_connector_forward # TODO ltx: patched diffusers connectors padding to fix audio token order (upstream #13564 regression)
+
+
+def apply_lora_patch():
+    try:
+        from diffusers.loaders.lora_pipeline import LTX2LoraLoaderMixin
+    except ImportError:
+        return
+    original = LTX2LoraLoaderMixin.lora_state_dict.__func__
+
+    @functools.wraps(original)
+    def lora_state_dict(cls, *args, **kwargs): # TODO ltx: diffusers routes 2.5 stage-2 lora connector keys into the transformer namespace
+        loaded = original(cls, *args, **kwargs)
+        if isinstance(loaded, tuple):
+            return (reroute_connector_keys(loaded[0]), *loaded[1:])
+        return reroute_connector_keys(loaded)
+
+    LTX2LoraLoaderMixin.lora_state_dict = classmethod(lora_state_dict)
+
+
+def apply_patch():
+    global PATCH_APPLIED # pylint: disable=global-statement
+    if PATCH_APPLIED:
+        return
+    apply_connectors_forward_patch()
+    apply_lora_patch()
+    PATCH_APPLIED = True

From f6607e89225eaa07d606e11ceee00b312998e13e Mon Sep 17 00:00:00 2001
From: CalamitousFelicitousness <iowasovereign@gmail.com>
Date: Wed, 12 Aug 2026 02:24:41 +0100
Subject: [PATCH 9/9] fix(hub): install the token header hijack at startup

Implicit tokens are disabled process-wide, so the hijack that re-attaches the
token explicitly is what keeps gated repos reachable. It was installed by the
checkpoint loader and the caption paths, so loading a video model into a fresh
session sent unauthenticated requests and gated repos answered 401.
---
 modules/models_hf.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/modules/models_hf.py b/modules/models_hf.py
index f6d311b8c..64f4f5b80 100644
--- a/modules/models_hf.py
+++ b/modules/models_hf.py
@@ -1,6 +1,7 @@
 import os
 import time
 import gradio as gr
+from modules import sd_hijack_hfhub
 from modules.logger import log
 from modules.shared import opts
 
@@ -37,6 +38,10 @@ def hf_init():
         obfuscated_token = 'hf_...' + opts.huggingface_token[-4:]
     log.info(f'Huggingface: transfer={opts.hf_transfer_mode} parallel={opts.sd_parallel_load} direct={opts.diffusers_to_gpu} token="{obfuscated_token}" cache="{opts.hfcache_dir}"')
 
+    # the disable flag above drops the token from token=None requests, which is what diffusers and
+    # transformers send; the hijack re-adds it explicitly and has to precede the first download
+    sd_hijack_hfhub.init_hijack()
+
 
 def hf_check_cache():
     t0 = time.time()