[bugfix] Use legacy LTX2 behavior for SSIM compatibility

Author: Davids048

fastvideo/fastvideo_args.py

@@ -206,6 +206,8 @@ class FastVideoArgs:
     ltx2_refine_add_noise: bool = True
     ltx2_refine_noise_path: str | None = None
     ltx2_refine_audio_noise_path: str | None = None
+    ltx2_legacy_native_noise_order: bool = False
+    ltx2_use_distilled_sigmas: bool = True
 
     # model paths for correct deallocation
     model_paths: dict[str, str] = field(default_factory=dict)

fastvideo/pipelines/basic/ltx2/stages/ltx2_denoising.py

@@ -243,7 +243,8 @@ def forward(
             logger.info("[LTX2] Using override sigma schedule, %s", self.sigmas_override)
         else:
             # Use distilled hardcoded schedule (or subsets) when enabled.
-            use_distilled_sigmas = os.getenv("LTX2_USE_DISTILLED_SIGMAS", "1") == "1"
+            use_distilled_sigmas = (fastvideo_args.ltx2_use_distilled_sigmas
+                                    and os.getenv("LTX2_USE_DISTILLED_SIGMAS", "1") == "1")
             max_distilled_steps = len(DISTILLED_SIGMA_VALUES) - 1
             if use_distilled_sigmas and num_inference_steps <= max_distilled_steps:
                 sigmas, distilled_indices = _distilled_subset_sigmas(

fastvideo/pipelines/basic/ltx2/stages/ltx2_latent_preparation.py

@@ -160,6 +160,14 @@ def forward(
                 loaded_latents = self._load_initial_latent(latent_path, device, dtype)
                 if loaded_latents is not None:
                     latents = loaded_latents
+                elif fastvideo_args.ltx2_legacy_native_noise_order:
+                    latents = randn_tensor(
+                        shape,
+                        generator=generator,
+                        device=device,
+                        dtype=dtype,
+                    )
+                    self._save_initial_latent(latent_path, latents)
                 else:
                     latents = _randn_ltx2_video_latents(
                         shape=shape,
@@ -170,13 +178,21 @@ def forward(
                     )
                     self._save_initial_latent(latent_path, latents)
             else:
-                latents = _randn_ltx2_video_latents(
-                    shape=shape,
-                    transformer=self.transformer,
-                    generator=generator,
-                    device=device,
-                    dtype=dtype,
-                )
+                if fastvideo_args.ltx2_legacy_native_noise_order:
+                    latents = randn_tensor(
+                        shape,
+                        generator=generator,
+                        device=device,
+                        dtype=dtype,
+                    )
+                else:
+                    latents = _randn_ltx2_video_latents(
+                        shape=shape,
+                        transformer=self.transformer,
+                        generator=generator,
+                        device=device,
+                        dtype=dtype,
+                    )
         else:
             latents = latents.to(device)
 

fastvideo/tests/ssim/test_ltx2_similarity.py

@@ -123,5 +123,10 @@ def test_ltx2_distilled_inference_similarity(
         full_quality_params_map=FULL_QUALITY_LTX2_DISTILLED_MODEL_TO_PARAMS,
         slice_cosine_threshold=SLICE_COSINE_DISTANCE_THRESHOLD,
         full_cosine_threshold=FULL_COSINE_DISTANCE_THRESHOLD,
+        init_kwargs_override={
+            "dit_cpu_offload": True,
+            "ltx2_legacy_native_noise_order": True,
+            "ltx2_use_distilled_sigmas": False,
+        },
         generation_kwargs_override=LTX2_DISTILLED_REFERENCE_GUIDANCE_OVERRIDES,
     )