fix latte1-t2v

Signed-off-by: Vladimir Mandic <mandic00@live.com>

TODO.md

@@ -7,9 +7,9 @@ Main ToDo list can be found at [GitHub projects](https://github.com/users/vladma
 ### Issues/Limitations
 
 - Video: Hunyuan Video I2V: requires `transformers==4.47.1` <https://github.com/huggingface/diffusers/issues/11118>  
-- Video: Latte 1 T2V: dtype mismatch <https://github.com/huggingface/diffusers/issues/11137>  
 - Video: CogVideoX 1.5 5B T2V/I2V: all-gray output  
 - Video: Allegro T2V: all-gray output
+- Video: Latte1 T2V: garbage output
 
 ## Future Candidates
 

installer.py

@@ -538,7 +538,7 @@ def check_diffusers():
     t_start = time.time()
     if args.skip_all or args.skip_git or args.experimental:
         return
-    sha = '617c208bb4cc68fe4518164fee7cbdf5aa44ff78' # diffusers commit hash
+    sha = '75d7e5cc459f66a53652445d5b281054b297680d' # diffusers commit hash
     pkg = pkg_resources.working_set.by_key.get('diffusers', None)
     minor = int(pkg.version.split('.')[1] if pkg is not None else 0)
     cur = opts.get('diffusers_version', '') if minor > 0 else ''

modules/ui_video.py

@@ -14,6 +14,14 @@ def engine_change(engine):
     return gr.update(choices=found, value=found[0] if len(found) > 0 else None)
 
 
+def get_selected(engine, model):
+    found = [model.name for model in models_def.models.get(engine, [])]
+    if len(models_def.models[engine]) > 0 and len(found) > 0:
+        selected = [m for m in models_def.models[engine] if m.name == model][0]
+        return selected
+    return None
+
+
 def model_change(engine, model):
     debug(f'Video change: engine="{engine}" model="{model}"')
     found = [model.name for model in models_def.models.get(engine, [])]
@@ -23,8 +31,7 @@ def model_change(engine, model):
 
 def model_load(engine, model):
     debug(f'Video load: engine="{engine}" model="{model}"')
-    found = [model.name for model in models_def.models.get(engine, [])]
-    selected = [m for m in models_def.models[engine] if m.name == model][0] if len(found) > 0 else None
+    selected = get_selected(engine, model)
     yield f'Video model loading: {selected.name}'
     if selected:
         if 'None' in selected.name:
@@ -43,8 +50,7 @@ def model_load(engine, model):
 def run_video(*args):
     engine, model = args[2], args[3]
     debug(f'Video run: engine="{engine}" model="{model}"')
-    found = [model.name for model in models_def.models.get(engine, [])]
-    selected = [m for m in models_def.models[engine] if m.name == model][0] if len(found) > 0 else None
+    selected = get_selected(engine, model)
     if not selected or engine is None or model is None or engine == 'None' or model == 'None':
         return video_utils.queue_err('model not selected')
     debug(f'Video run: {str(selected)}')

modules/video_models/video_overrides.py

@@ -27,17 +27,19 @@ def set_overrides(p: processing.StableDiffusionProcessingVideo, selected: Model)
         shared.sd_model.vae.enable_tiling()
     # Latte
     if selected.name == 'Latte 1 T2V':
-        p.task_args['enable_temporal_attentions'] = False
-        p.task_args['video_length'] = p.frames
+        p.task_args['enable_temporal_attentions'] = True
+        p.task_args['video_length'] = 16 * (max(p.frames // 16, 1))
     # LTX
     if cls == 'LTXImageToVideoPipeline' or cls == 'LTXConditionPipeline':
         p.task_args['generator'] = None
     if cls == 'LTXConditionPipeline':
         p.task_args['strength'] = p.denoising_strength
+    # WAN
     if 'Wan' in cls:
         p.task_args['width'] = 16 * (p.width // 16)
         p.task_args['height'] = 16 * (p.height // 16)
-        p.frames = 4 * (p.frames // 4) + 1
+        p.frames = 4 * (max(p.frames // 4, 1)) + 1
+    # LTX
     if 'LTX' in cls:
         p.task_args['width'] = 32 * (p.width // 32)
         p.task_args['height'] = 32 * (p.height // 32)