minimax redo video/audio shift

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

CHANGELOG.md

@@ -49,6 +49,7 @@ Plus inevitable bug-fixes...
     can significantly improve lora quality when using multiple loras at once  
   - per-block strength
   - native support for **MiniMax**
+    see [MiniMax Turbo LoRA collection](https://huggingface.co/vladmandic/MiniMax-H3-Turbo-LoRA) for LoRAs and examples  
 - **DLSS**
   - add DLSS support for: *NeuralRender, SuperSample and FrameGen*  
     dlls 5 caused quite a stir, but combined with generative ai it becomes a nice tool  
@@ -107,6 +108,7 @@ Plus inevitable bug-fixes...
     *note*: this may break compatibility with some legacy packages, so report any finidings  
 - **Other**
   - video preview: TAESD support for **MiniMax**
+  - **minimax** video and audio shift are now dependent on steps instead of static
   - support `xai grok` for prompt enhance workflows  
     *note*: requires grok api key  
   - remove `/redocs` as `/docs` are primary api docs  

modules/minimax/minimax_ui.py

@@ -29,8 +29,8 @@ def create_ui(prompt, _negative, styles, overrides, script_inputs, mp4_fps, mp4_
                     steps = gr.Slider(minimum=2, maximum=100, step=1, label="MiniMax steps", elem_id='minimax_steps', value=30)
                     frames = gr.Slider(label='MiniMax frames', minimum=22, maximum=362, step=17, value=124, elem_id='minimax_frames')
                 with gr.Row():
-                    video_shift = gr.Slider(minimum=8.0, maximum=16.0, step=0.1, label="MiniMax video shift", elem_id='minimax_video_shift', value=12)
-                    audio_shift = gr.Slider(minimum=1.5, maximum=6.0, step=0.1, label="MiniMax audio shift", elem_id='minimax_audio_shift', value=3)
+                    video_shift = gr.Slider(minimum=0.05, maximum=0.95, step=0.05, value=0.40, label="MiniMax video shift", elem_id='minimax_video_shift')
+                    audio_shift = gr.Slider(minimum=0.05, maximum=0.95, step=0.05, value=0.15, label="MiniMax audio shift", elem_id='minimax_audio_shift')
                 with gr.Row():
                     seed = gr.Number(label='Seed', value=-1, elem_id='minimax_seed', container=True)
                     random_seed = ToolButton(ui_symbols.random, elem_id='minimax_seed_random')

modules/minimax/minimax_video.py

@@ -145,8 +145,8 @@ def generate(task_id, _ui_state,
                 ops=['video'],
             )
             video_minimax.apply_overrides(p, shared.sd_model, still=False, audio=enable_audio, preview=enable_preview)
-            video_minimax.set_sampler_shift(shared.sd_model, video_shift=video_shift, audio_shift=audio_shift)
-            log.debug(f'Video: engine="{engine}" model="{model}" workflow={workflow} cls={shared.sd_model.__class__.__name__} audio={enable_audio} preview={enable_preview} shift={video_shift}:{audio_shift} kwargs={p.task_args}')
+            video_minimax.set_sampler_shift(shared.sd_model, steps=steps, video_shift=video_shift, audio_shift=audio_shift)
+            log.debug(f'Video: engine="{engine}" model="{model}" workflow={workflow} cls={shared.sd_model.__class__.__name__} audio={enable_audio} preview={enable_preview} kwargs={p.task_args}')
             processing.fix_seed(p)
             p.ops.append('video')
             p.scripts = scripts_manager.scripts_video

modules/video_models/video_minimax.py

@@ -91,8 +91,25 @@ def set_audio(pipe, enabled: bool):
         log.debug(f'Pipeline: cls={pipe.__class__.__name__} audio=disabled')
 
 
-def set_sampler_shift(pipe, video_shift: float = 12.0, audio_shift: float = 3.0):
-    if getattr(pipe, 'scheduler', None) is not None and getattr(pipe.scheduler, 'config', None) is not None:
-        pipe.scheduler.config.shift = video_shift
-    if getattr(pipe, 'audio_scheduler', None) is not None and getattr(pipe.audio_scheduler, 'config', None) is not None:
-        pipe.audio_scheduler.config.shift = audio_shift
+def calculate_video_shift(steps: int, value: float = 0.40, max_shift: float = 16.0) -> int:
+    value = max(0.05, min(0.95, value))
+    return min(max_shift, round((value * steps) + 0.5))
+
+
+def calculate_audio_shift(steps: int, value: float = 0.15, max_shift: float = 6.0) -> int:
+    value = max(0.05, min(0.95, value))
+    return min(max_shift, round((value * steps) + 0.5))
+
+
+def set_sampler_shift(pipe, steps: int, video_shift: float = 12.0, audio_shift: float = 3.0):
+    import diffusers
+    if getattr(pipe, 'scheduler', None) is None or getattr(pipe.scheduler, 'config', None) is None:
+        log.warning(f'Pipeline: cls={pipe.__class__.__name__} scheduler is missing')
+        return
+    video_calc_shift = calculate_video_shift(steps=steps, value=video_shift)
+    audio_calc_shift = calculate_audio_shift(steps=steps, value=audio_shift)
+    dct_video = { 'shift': video_shift, 'steps': video_calc_shift }
+    dct_audio = { 'shift': audio_shift, 'steps': audio_calc_shift }
+    pipe.scheduler = diffusers.MiniMaxH3Scheduler(shift=video_calc_shift)
+    pipe.audio_scheduler.config.shift = diffusers.MiniMaxH3Scheduler(shift=audio_calc_shift)
+    log.debug(f'Pipeline: scheduler={pipe.scheduler.__class__.__name__} video={dct_video} audio={dct_audio} ')