fix(video): count minimax steps as transformer evaluations

MiniMaxH3Scheduler counts the terminal sigma in num_inference_steps, so
Steps N ran N-1 evaluations while every other model runs N. The shim
hands the scheduler p.steps + 1, the slider starts at 1, and the PDD
pin records the evaluation count while passing the scheduler its grid
argument. Metadata written before this change counted grid points.

modules/lora/network_pdd.py

@@ -285,7 +285,7 @@ def reconcile(pipe, loaded, components):
 
 
 def pin(p, model):
-    """Hold a generation on the distilled step count and shipped schedule while heads are installed; returns the pinned step count or None."""
+    """Hold a generation on the distilled evaluation count and shipped schedule while heads are installed; returns the scheduler step argument or None."""
     state = getattr(model, 'sdnext_pdd', None)
     if state is None:
         return None
@@ -297,13 +297,13 @@ def pin(p, model):
             scheduler.set_shift(scheduler.config['shift'])
             shifts[name] = scheduler.config['shift']
     requested = p.steps
-    p.steps = state.steps
+    p.steps = state.heads.nfe # what the user-facing step count means: transformer evaluations
     if getattr(p, 'task_args', None) is not None:
-        p.task_args['num_inference_steps'] = state.steps
+        p.task_args['num_inference_steps'] = state.steps # the scheduler argument that yields that many grid intervals
     if getattr(model, 'num_timesteps', None) is not None:
         model.num_timesteps = state.heads.nfe # the progress total counts transformer evaluations
     extra = getattr(p, 'extra_generation_params', None)
     if extra is not None:
         extra.update({state.spec.shift_keys[name]: shift for name, shift in shifts.items() if name in state.spec.shift_keys})
-    log.info(f'Network: type=PDD name="{state.name}" steps={state.steps} requested={requested} nfe={state.heads.nfe} shift={shifts}')
+    log.info(f'Network: type=PDD name="{state.name}" steps={state.heads.nfe} requested={requested} grid_steps={state.steps} shift={shifts}')
     return state.steps

modules/minimax/minimax_ui.py

@@ -26,7 +26,7 @@ def create_ui(prompt, _negative, styles, overrides, script_inputs, mp4_fps, mp4_
                     width, height = ui_sections.create_resolution_inputs('minimax', default_width=1024, default_height=576, step=32)
                     btn_detect_image_size = ToolButton(value=ui_symbols.detect, elem_id="minimax_resize_detect_size")
                 with gr.Row():
-                    steps = gr.Slider(minimum=2, maximum=100, step=1, label="MiniMax steps", elem_id='minimax_steps', value=30)
+                    steps = gr.Slider(minimum=1, 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=0.5, maximum=20.0, step=0.1, value=12.0, label="MiniMax video shift", elem_id='minimax_video_shift')

modules/video_models/video_minimax.py

@@ -29,9 +29,9 @@ def apply_overrides(p, pipe, still: bool = False, audio: bool = True, preview: b
         log.debug(f'Pipeline: cls={pipe.__class__.__name__} frames requested={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
+    p.steps = max(1, p.steps) # transformer evaluations, as on every other model
+    p.task_args['num_inference_steps'] = p.steps + 1 # the scheduler counts the terminal sigma as a grid point
+    pipe.num_timesteps = p.steps # 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

test/test-minimax-shift.py

@@ -6,7 +6,7 @@ Offline unit tests for the MiniMax schedule shift in modules.video_models.video_
 - ``set_sampler_shift`` writes scheduler, default_scheduler and audio_scheduler and keys the result for infotext
 - a request without values resets what the previous request set
 - the Default sampler restore, a deepcopy of default_scheduler, carries the shift into the sigma grid
-- ``apply_overrides`` records the applied values on the processing object
+- ``apply_overrides`` records the applied values on the processing object and hands the scheduler one grid point more than the step count
 
 No running server required.
 
@@ -188,6 +188,7 @@ def test_apply_overrides_records_the_applied_values():
     assert p.extra_generation_params == {'Video shift': 6.0, 'Audio shift': AUDIO_SHIFT}, f'recorded={p.extra_generation_params}'
     assert shifts(pipe) == (6.0, 6.0, AUDIO_SHIFT), f'shifts={shifts(pipe)}'
     assert p.sampler_name == 'Default'
+    assert p.steps == STEPS and p.task_args['num_inference_steps'] == STEPS + 1 and pipe.num_timesteps == STEPS, f'steps={p.steps} grid_steps={p.task_args["num_inference_steps"]} total={pipe.num_timesteps}'
 
 
 def test_apply_overrides_without_values_uses_the_shipped_schedule():

test/test-pdd.py

@@ -335,7 +335,7 @@ def test_pin_overrides_steps_and_shift():
     pipe.audio_scheduler.set_shift(2.0)
     p = types.SimpleNamespace(steps=30, task_args={'num_inference_steps': 30}, extra_generation_params={'Video shift': 4.0, 'Audio shift': 2.0})
     assert network_pdd.pin(p, pipe) == 9
-    assert p.steps == 9 and p.task_args['num_inference_steps'] == 9
+    assert p.steps == 8 and p.task_args['num_inference_steps'] == 9, f'steps={p.steps} grid_steps={p.task_args["num_inference_steps"]}'
     assert pipe.num_timesteps == 8
     assert pipe.scheduler.shift == 12.0 and pipe.audio_scheduler.shift == 3.0
     assert p.extra_generation_params == {'Video shift': 12.0, 'Audio shift': 3.0}, f'recorded={p.extra_generation_params}'

ui/locale/locale_en.json

@@ -949,6 +949,7 @@
     {"id":"minimax_video_shift","label":"MiniMax video shift","localized":"","hint":"Exponential shift of the video sigma schedule, <code>sigma' = s*sigma / (1 + (s-1)*sigma)</code>. Values above 1 move the step grid toward full noise, values below 1 toward the clean end. The value is absolute and does not scale with the step count.<br><br>Default is <b>12</b>, the value the model ships with. Distilled LoRAs run at the shift they were trained with: <b>12</b> for the 544p <i>lightx2v</i> files, <b>6</b> for their 768p files. Parallel decoding (PDD) LoRAs pin the shipped value.<br><br>Recorded in the output metadata as <b>Video shift</b>.","ui":"video"},
     {"id":"minimax_audio_shift","label":"MiniMax audio shift","localized":"","hint":"Exponential shift of the audio sigma schedule. The audio rows are denoised on this schedule inside the joint pass, so the value applies with audio output disabled too.<br><br>Default is <b>3</b>, the value the model ships with; the published turbo LoRAs keep it. Parallel decoding (PDD) LoRAs pin the shipped value.<br><br>Recorded in the output metadata as <b>Audio shift</b>.","ui":"video"},
     {"id":"","label":"MiniMax Frames","localized":"","hint":"MiniMax is optimized to generate 5-15sec videos at 24 FPS","ui":"video"},
+    {"id":"minimax_steps","label":"MiniMax steps","localized":"","hint":"Number of transformer evaluations, counted the same way as on every other model. Distilled LoRAs run at the count in their name: a 4-step file at <b>4</b>, an 8-step file at <b>8</b>. Parallel decoding (PDD) LoRAs pin their own count.<br><br>Default is <b>30</b>.","ui":"video"},
         {"id":"","label":"Model Info","localized":"","hint":"","ui":"component-8779"},
     {"id":"","label":"Model pipeline","localized":"","hint":"If autodetect does not detect model automatically, select model type before loading a model","ui":"settings_sd"},
     {"id":"","label":"Model auto-load on start","localized":"","hint":"","ui":"settings_sd"},