Automated PR - 2026-03-04

2026-03-04 19:34:46 +00:00
parent 28c3c73fe5
commit 822ce3c4b1
73 changed files with 4984 additions and 1220 deletions
@@ -1,12 +1,17 @@
+import logging
+from dataclasses import dataclass, field, replace
+
+from safetensors import safe_open
+
+from ltx_core.components.guiders import MultiModalGuiderParams
+from ltx_core.types import SpatioTemporalScaleFactors
+
 # =============================================================================
 # Diffusion Schedule
 # =============================================================================

 # Noise schedule for the distilled pipeline. These sigma values control noise
 # levels at each denoising step and were tuned to match the distillation process.
-from ltx_core.components.guiders import MultiModalGuiderParams
-from ltx_core.types import SpatioTemporalScaleFactors
-
 DISTILLED_SIGMA_VALUES = [1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0]

 # Reduced schedule for super-resolution stage 2 (subset of distilled values)
@@ -14,64 +19,88 @@ STAGE_2_DISTILLED_SIGMA_VALUES = [0.909375, 0.725, 0.421875, 0.0]


 # =============================================================================
-# Video Generation Defaults
+# Pipeline Parameters
 # =============================================================================

-DEFAULT_SEED = 10
-DEFAULT_1_STAGE_HEIGHT = 512
-DEFAULT_1_STAGE_WIDTH = 768
-DEFAULT_2_STAGE_HEIGHT = DEFAULT_1_STAGE_HEIGHT * 2
-DEFAULT_2_STAGE_WIDTH = DEFAULT_1_STAGE_WIDTH * 2
-DEFAULT_NUM_FRAMES = 121
-DEFAULT_FRAME_RATE = 24.0
-DEFAULT_NUM_INFERENCE_STEPS = 40
-DEFAULT_VIDEO_GUIDER_PARAMS = MultiModalGuiderParams(
-    cfg_scale=3.0,
-    stg_scale=1.0,
-    rescale_scale=0.7,
-    modality_scale=3.0,
-    skip_step=0,
-    stg_blocks=[29],
+
+@dataclass(frozen=True)
+class PipelineParams:
+    seed: int = 10
+    stage_1_height: int = 512
+    stage_1_width: int = 768
+    num_frames: int = 121
+    frame_rate: float = 24.0
+    num_inference_steps: int = 40
+    video_guider_params: MultiModalGuiderParams = field(
+        default_factory=lambda: MultiModalGuiderParams(
+            cfg_scale=3.0,
+            stg_scale=1.0,
+            rescale_scale=0.7,
+            modality_scale=3.0,
+            skip_step=0,
+            stg_blocks=[29],
+        )
+    )
+    audio_guider_params: MultiModalGuiderParams = field(
+        default_factory=lambda: MultiModalGuiderParams(
+            cfg_scale=7.0,
+            stg_scale=1.0,
+            rescale_scale=0.7,
+            modality_scale=3.0,
+            skip_step=0,
+            stg_blocks=[29],
+        )
+    )
+
+    @property
+    def stage_2_height(self) -> int:
+        return int(self.stage_1_height * 2)
+
+    @property
+    def stage_2_width(self) -> int:
+        return int(self.stage_1_width * 2)
+
+
+# Default params for LTX-2.0 non-distilled models. These can be overridden by detecting from checkpoint metadata.
+LTX_2_PARAMS = PipelineParams()
+
+# Default params for LTX-2.3 non-distilled models. These override some of the LTX-2.0 defaults.
+LTX_2_3_PARAMS = replace(
+    LTX_2_PARAMS,
+    num_inference_steps=30,
+    video_guider_params=replace(LTX_2_PARAMS.video_guider_params, stg_blocks=[28]),
+    audio_guider_params=replace(LTX_2_PARAMS.audio_guider_params, stg_blocks=[28]),
 )

-
-# =============================================================================
-# Audio
-# =============================================================================
-
-DEFAULT_AUDIO_GUIDER_PARAMS = MultiModalGuiderParams(
-    cfg_scale=7.0,
-    stg_scale=1.0,
-    rescale_scale=0.7,
-    modality_scale=3.0,
-    skip_step=0,
-    stg_blocks=[29],
-)
-AUDIO_SAMPLE_RATE = 24000
-
-
-# =============================================================================
-# LoRA
-# =============================================================================
-
 DEFAULT_LORA_STRENGTH = 1.0
-
-
-# =============================================================================
-# Video VAE Architecture
-# =============================================================================
-
+DEFAULT_IMAGE_CRF = 33
 VIDEO_SCALE_FACTORS = SpatioTemporalScaleFactors.default()
 VIDEO_LATENT_CHANNELS = 128

+_LTX_2_3_MODEL_VERSION_PREFIX = "2.3"

-# =============================================================================
-# Image Preprocessing
-# =============================================================================

-# CRF (Constant Rate Factor) for H.264 encoding used in image conditioning.
-# Lower = higher quality, 0 = lossless. This mimics compression artifacts.
-DEFAULT_IMAGE_CRF = 33
+def detect_params(checkpoint_path: str) -> PipelineParams:
+    """Detect pipeline params from checkpoint metadata.
+    Reads the ``model_version`` field from the safetensors config metadata.
+    Returns ``LTX_2_3_PARAMS`` when the version starts with "2.3",
+    otherwise falls back to ``LTX_2_PARAMS``.
+    """
+    logger = logging.getLogger(__name__)
+
+    try:
+        with safe_open(checkpoint_path, framework="pt") as f:
+            metadata = f.metadata() or {}
+        version = metadata.get("model_version", "")
+    except Exception:
+        logger.warning("Could not read checkpoint metadata from %s, using LTX-2 defaults", checkpoint_path)
+        return LTX_2_PARAMS
+
+    if version.startswith(_LTX_2_3_MODEL_VERSION_PREFIX):
+        return LTX_2_3_PARAMS
+
+    logger.info("Using LTX_2_PARAMS for checkpoint (version=%s)", version or "unknown")
+    return LTX_2_PARAMS


 # =============================================================================