missing file

Detail:

- Merged LTX2 sampling behavior into the global registry and removed the ambiguous local converted-path default.
- Replaced single LTX2 sampling mapping with explicit model-ID mappings:
    - Lightricks/LTX-2 -> LTX2BaseSamplingParam
    - FastVideo/LTX2-base -> LTX2BaseSamplingParam
    - FastVideo/LTX2-Distilled-Diffusers -> LTX2DistilledSamplingParam
- Kept LTX2T2VConfig as the pipeline config for all explicitly mapped LTX2 IDs.
- Removed implicit mapping for converted/ltx2_diffusers to avoid guessing base vs distilled for user-local
  conversions.
- Added focused local tests at tests/local_tests/test_ltx2_registry.py for:
    - exact base/distilled sampling resolution,
    - pipeline config resolution,
    - no fallback behavior for ambiguous local converted paths.

Assumptions:

- Canonical rename/ID intent:
    - “base” names map to LTX2BaseSamplingParam.
    - “Distilled” names map to LTX2DistilledSamplingParam.
- converted/ltx2_diffusers is intentionally ambiguous across users and must not be auto-assigned.
- Unknown/non-canonical names containing “LTX”/“LTX2” (but not matching explicit registered IDs) should not auto-
  resolve to base or distilled.
    - Sampling resolver returns None (caller falls back to generic defaults/user overrides).
    - Pipeline config lookup raises a “No match found” error.

Notes:

- This change prioritizes explicitness over convenience: only predetermined, canonical model IDs get LTX2-specific
  defaults; everything else requires user intent.

CLAUDE.md

@@ -1 +1 @@
-@AGENTS.md
+@AGENTS.md

examples/inference/basic/basic_ltx2.py

@@ -1,5 +1,6 @@
 from fastvideo import VideoGenerator
 
+
 PROMPT = (
     "A warm sunny backyard. The camera starts in a tight cinematic close-up "
     "of a woman and a man in their 30s, facing each other with serious "
@@ -16,16 +17,20 @@ PROMPT = (
 
 
 def main() -> None:
+    # Uses FastVideo default sampling settings for LTX2 base.
     generator = VideoGenerator.from_pretrained(
-        "FastVideo/LTX2-Distilled-Diffusers",
+        "Davids048/LTX2-Base-Diffusers",
         num_gpus=1,
     )
 
-    output_path = "outputs_video/ltx2_basic/output_ltx2_distilled_t2v.mp4"
+    output_path = "outputs_video/ltx2_basic/output_ltx2_base_t2v_1088_1920_1.1.mp4"
     generator.generate_video(
         prompt=PROMPT,
         output_path=output_path,
         save_video=True,
+        num_frames=121,
+        height=1088,
+        width=1920,
     )
     generator.shutdown()
 

examples/inference/basic/basic_ltx2_base.py

@@ -1,54 +0,0 @@
-from fastvideo import VideoGenerator
-
-
-PROMPT = (
-    "A warm sunny backyard. The camera starts in a tight cinematic close-up "
-    "of a woman and a man in their 30s, facing each other with serious "
-    "expressions. The woman, emotional and dramatic, says softly, \"That's "
-    "it... Dad's lost it. And we've lost Dad.\" The man exhales, slightly "
-    "annoyed: \"Stop being so dramatic, Jess.\" A beat. He glances aside, "
-    "then mutters defensively, \"He's just having fun.\" The camera slowly "
-    "pans right, revealing the grandfather in the garden wearing enormous "
-    "butterfly wings, waving his arms in the air like he's trying to take "
-    "off. He shouts, \"Wheeeew!\" as he flaps his wings with full commitment. "
-    "The woman covers her face, on the verge of tears. The tone is deadpan, "
-    "absurd, and quietly tragic."
-)
-
-official_negative_prompt = (
-    "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, "
-    "grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, "
-    "deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, "
-    "wrong hand count, artifacts around text, inconsistent perspective, camera shake, incorrect depth of "
-    "field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent "
-    "lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny "
-    "valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, wrong gaze direction, "
-    "mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, "
-    "off-sync audio, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward "
-    "pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, flat lighting, "
-    "inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts."
-)
-def main() -> None:
-    # Uses FastVideo default sampling settings for LTX2 base.
-    generator = VideoGenerator.from_pretrained(
-        "Davids048/LTX2-Base-Diffusers",
-        num_gpus=1,
-    )
-
-    output_path = "outputs_video/ltx2_basic/output_ltx2_base_t2v_1088_1920_1.0.mp4"
-    generator.generate_video(
-        prompt=PROMPT,
-        output_path=output_path,
-        save_video=True,
-        num_frames=121,
-        height=1088,
-        width=1920,
-        fps=24,
-        negative_prompt=official_negative_prompt,
-        guidance_scale=3.0,
-    )
-    generator.shutdown()
-
-
-if __name__ == "__main__":
-    main()

examples/inference/basic/basic_ltx2_distilled.py

@@ -0,0 +1,34 @@
+from fastvideo import VideoGenerator
+
+PROMPT = (
+    "A warm sunny backyard. The camera starts in a tight cinematic close-up "
+    "of a woman and a man in their 30s, facing each other with serious "
+    "expressions. The woman, emotional and dramatic, says softly, \"That's "
+    "it... Dad's lost it. And we've lost Dad.\" The man exhales, slightly "
+    "annoyed: \"Stop being so dramatic, Jess.\" A beat. He glances aside, "
+    "then mutters defensively, \"He's just having fun.\" The camera slowly "
+    "pans right, revealing the grandfather in the garden wearing enormous "
+    "butterfly wings, waving his arms in the air like he's trying to take "
+    "off. He shouts, \"Wheeeew!\" as he flaps his wings with full commitment. "
+    "The woman covers her face, on the verge of tears. The tone is deadpan, "
+    "absurd, and quietly tragic."
+)
+
+
+def main() -> None:
+    generator = VideoGenerator.from_pretrained(
+        "FastVideo/LTX2-Distilled-Diffusers",
+        num_gpus=1,
+    )
+
+    output_path = "outputs_video/ltx2_basic/output_ltx2_distilled_t2v.mp4"
+    generator.generate_video(
+        prompt=PROMPT,
+        output_path=output_path,
+        save_video=True,
+    )
+    generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

fastvideo/registry.py

@@ -249,8 +249,8 @@ def _register_configs() -> None:
             "FastVideo/LTX2-base",
         ],
         model_detectors=[
-            lambda path: ("ltx2" in path.lower() or "ltx-2" in path.lower())
+            lambda path: ("ltx2" in path.lower() or "ltx-2" in path.lower()) and
-            and "distilled" not in path.lower(),
+            "distilled" not in path.lower(),
         ],
     )
     # LTX-2 (distilled)
@@ -261,7 +261,8 @@ def _register_configs() -> None:
             "FastVideo/LTX2-Distilled-Diffusers",
         ],
         model_detectors=[
-            lambda path: ("ltx2" in path.lower() or "ltx-2" in path.lower()) and "distilled" in path.lower(),
+            lambda path: ("ltx2" in path.lower() or "ltx-2" in path.lower()) and
+            "distilled" in path.lower(),
         ],
     )
 

fastvideo/training/ltx2_training_pipeline.py

@@ -280,18 +280,19 @@ class LTX2TrainingPipeline(TrainingPipeline):
 
         conditioning_mask = None
         first_frame_p = self.training_args.ltx2_first_frame_conditioning_p
-        if first_frame_p > 0:
+        if (first_frame_p > 0 and torch.rand(
-            if torch.rand(1, generator=self.noise_random_generator).item(
+                1,
-            ) < first_frame_p:
+                generator=self.noise_random_generator,
-                conditioning_mask = torch.zeros(
+        ).item() < first_frame_p):
-                    (batch_size, 1, latents.shape[2], latents.shape[3],
+            conditioning_mask = torch.zeros(
-                     latents.shape[4]),
+                (batch_size, 1, latents.shape[2], latents.shape[3],
-                    dtype=torch.bool,
+                 latents.shape[4]),
-                    device=latents.device,
+                dtype=torch.bool,
-                )
+                device=latents.device,
-                conditioning_mask[:, :, 0:1] = True
+            )
-                noisy_model_input = torch.where(conditioning_mask, latents,
+            conditioning_mask[:, :, 0:1] = True
-                                                noisy_model_input)
+            noisy_model_input = torch.where(conditioning_mask, latents,
+                                            noisy_model_input)
 
         if conditioning_mask is None:
             mask_patch = torch.zeros(
@@ -398,9 +399,9 @@ class LTX2TrainingPipeline(TrainingPipeline):
         with self.tracker.timed("timing/forward_backward"), set_forward_context(
                 current_timestep=training_batch.current_timestep,
                 attn_metadata=training_batch.attn_metadata):
-            with torch.autocast("cuda", dtype=training_batch.latents.dtype):
+            with torch.autocast("cuda", dtype=training_batch.latents.dtype
-                with torch.autograd.set_detect_anomaly(True):
+                                ), torch.autograd.set_detect_anomaly(True):
-                    outputs = self.transformer(**input_kwargs)
+                outputs = self.transformer(**input_kwargs)
             if isinstance(outputs, tuple):
                 video_denoised, audio_denoised = outputs
             else:
@@ -422,7 +423,7 @@ class LTX2TrainingPipeline(TrainingPipeline):
                 audio_loss = (audio_pred_velocity.float() -
                               audio_target.float())**2
                 loss = loss + audio_loss.mean()
-                logger.info("Audio loss: {}".format(audio_loss.mean()))
+                logger.info("Audio loss: %s", audio_loss.mean())
             else:
                 logger.warning("Audio denoised is None")
 

fastvideo/training/training_pipeline.py

@@ -888,59 +888,61 @@ class TrainingPipeline(LoRAPipeline, ABC):
 
             # Only sp_group leaders (rank_in_sp_group == 0) need to send their
             # results to global rank 0
-            if self.rank_in_sp_group == 0:
+            if self.rank_in_sp_group == 0 and self.global_rank == 0:
-                if self.global_rank == 0:
+                # Global rank 0 collects results from all sp_group leaders
-                    # Global rank 0 collects results from all sp_group leaders
+                all_videos = step_videos  # Start with own results
-                    all_videos = step_videos  # Start with own results
+                all_captions = step_captions
-                    all_captions = step_captions
 
-                    # Receive from other sp_group leaders
+                # Receive from other sp_group leaders
-                    for sp_group_idx in range(1, num_sp_groups):
+                for sp_group_idx in range(1, num_sp_groups):
-                        src_rank = sp_group_idx * self.sp_world_size  # Global rank of other sp_group leaders
+                    src_rank = sp_group_idx * self.sp_world_size  # Global rank of other sp_group leaders
-                        recv_videos = world_group.recv_object(src=src_rank)
+                    recv_videos = world_group.recv_object(src=src_rank)
-                        recv_captions = world_group.recv_object(src=src_rank)
+                    recv_captions = world_group.recv_object(src=src_rank)
-                        all_videos.extend(recv_videos)
+                    all_videos.extend(recv_videos)
-                        all_captions.extend(recv_captions)
+                    all_captions.extend(recv_captions)
 
-                    video_filenames = []
+                video_filenames = []
-                    for i, (video, caption) in enumerate(
+                for i, (video, caption) in enumerate(
-                            zip(all_videos, all_captions, strict=True)):
+                        zip(all_videos, all_captions, strict=True)):
-                        os.makedirs(training_args.output_dir, exist_ok=True)
+                    os.makedirs(training_args.output_dir, exist_ok=True)
-                        filename = os.path.join(
+                    filename = os.path.join(
-                            training_args.output_dir,
+                        training_args.output_dir,
-                            f"validation_step_{global_step}_inference_steps_{num_inference_steps}_video_{i}.mp4"
+                        f"validation_step_{global_step}_inference_steps_{num_inference_steps}_video_{i}.mp4"
-                        )
+                    )
-                        imageio.mimsave(filename, video, fps=sampling_param.fps)
+                    imageio.mimsave(filename, video, fps=sampling_param.fps)
-                        # Mux audio if available
+                    # Mux audio if available
-                        audio = output_batch.extra.get("audio")
+                    audio = output_batch.extra.get("audio")
-                        audio_sample_rate = output_batch.extra.get(
+                    audio_sample_rate = output_batch.extra.get(
-                            "audio_sample_rate")
+                        "audio_sample_rate")
-                        if audio is not None and audio_sample_rate is not None:
+                    if (audio is not None and audio_sample_rate is not None
-                            if not self._mux_audio(filename, audio,
+                            and not self._mux_audio(
-                                                   audio_sample_rate):
+                                filename,
-                                logger.warning(
+                                audio,
-                                    "Audio mux failed for validation video %s; saved video without audio.",
+                                audio_sample_rate,
-                                    filename)
+                            )):
-                        video_filenames.append(filename)
+                        logger.warning(
+                            "Audio mux failed for validation video %s; saved video without audio.",
+                            filename)
+                    video_filenames.append(filename)
 
-                    artifacts = []
+                artifacts = []
-                    for filename, caption in zip(video_filenames,
+                for filename, caption in zip(video_filenames,
-                                                 all_captions,
+                                             all_captions,
-                                                 strict=True):
+                                             strict=True):
-                        video_artifact = self.tracker.video(filename,
+                    video_artifact = self.tracker.video(filename,
-                                                            caption=caption)
+                                                        caption=caption)
-                        if video_artifact is not None:
+                    if video_artifact is not None:
-                            artifacts.append(video_artifact)
+                        artifacts.append(video_artifact)
-                    if artifacts:
+                if artifacts:
-                        logs = {
+                    logs = {
-                            f"validation_videos_{num_inference_steps}_steps":
+                        f"validation_videos_{num_inference_steps}_steps":
-                            artifacts
+                        artifacts
-                        }
+                    }
-                        self.tracker.log_artifacts(logs, global_step)
+                    self.tracker.log_artifacts(logs, global_step)
-                else:
+            elif self.rank_in_sp_group == 0:
-                    # Other sp_group leaders send their results to global rank 0
+                # Other sp_group leaders send their results to global rank 0
-                    world_group.send_object(step_videos, dst=0)
+                world_group.send_object(step_videos, dst=0)
-                    world_group.send_object(step_captions, dst=0)
+                world_group.send_object(step_captions, dst=0)
 
         # Re-enable gradients for training
         training_args.inference_mode = False