Skip to content

armgit / LTX-2

Public Imported from GitHub
0

Official Python inference and LoRA trainer package for the LTX-2 audio–video generative model.

Automated PR - 2026-09-30

github-actions[bot]2d6e71c88be3
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 745395e..9087fd3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,11 @@
# Changelog
+## 1.4.1 - 2026-09-30
+
+### Changed
+
+- HDR video encoding streams decoded chunks directly into EXR and HLG outputs instead of retaining the full clip in memory.
+
## 1.4.0 - 2026-09-29
### Added
diff --git a/packages/ltx-core/pyproject.toml b/packages/ltx-core/pyproject.toml
index 106970b..0b6d888 100644
--- a/packages/ltx-core/pyproject.toml
+++ b/packages/ltx-core/pyproject.toml
@@ -1,6 +1,6 @@
[project]
name = "ltx-core"
-version = "1.4.0"
+version = "1.4.1"
description = "Core implementation of Lightricks' LTX-2 model"
readme = "README.md"
requires-python = ">=3.10"
diff --git a/packages/ltx-core/src/ltx_core/color/hlg.py b/packages/ltx-core/src/ltx_core/color/hlg.py
index 4660fde..5916f2b 100644
--- a/packages/ltx-core/src/ltx_core/color/hlg.py
+++ b/packages/ltx-core/src/ltx_core/color/hlg.py
@@ -196,9 +196,6 @@ def _hlg_encode_planes_threaded(
encoder: HlgPyAVEncoder,
plane_chunks: Iterator[tuple[np.ndarray, np.ndarray, np.ndarray]],
progress_total: int | None = None,
- *,
- audio: Audio | None = None,
- audio_stream: av.audio.AudioStream | None = None,
) -> int:
"""libx265-encode planar uint16 chunks on a background thread (1-slot queue)."""
chunk_queue: Queue[tuple[np.ndarray, np.ndarray, np.ndarray] | None] = Queue(maxsize=1)
@@ -223,9 +220,6 @@ def _hlg_encode_planes_threaded(
if error is None:
try:
encoder.flush()
- if audio is not None and audio_stream is not None:
- write_audio(encoder.container, audio_stream, audio)
- encoder.container.close()
except Exception as e:
error = e
else:
@@ -273,13 +267,27 @@ def encode_linear_hdr_frames_to_hlg_mp4( # noqa: PLR0913
progress: bool = False,
progress_total: int | None = None,
thread_count: int = 0,
- audio: Audio | None = None,
+ audio: Audio | Iterator[Audio] | None = None,
device: torch.device | None = None,
+ audio_sampling_rate: int | None = None,
) -> int:
"""Encode scene-linear HDR frames to BT.2020/HLG/10-bit HEVC mp4 via PyAV.
Consumes ``HxWx3`` / ``FxHxWx3`` chunks one at a time. Conversion runs on
- ``device``; the encoder thread only muxes planar uint16 into libx265.
+ ``device``; the encoder thread only muxes planar uint16 into libx265. An
+ audio iterator is pulled after all frame chunks have been consumed; pass
+ ``audio_sampling_rate`` so its AAC stream can be opened up front.
"""
+ if isinstance(audio, Audio):
+ validate_audio_waveform(audio)
+ if audio_sampling_rate is not None and audio_sampling_rate != audio.sampling_rate:
+ raise ValueError(
+ f"audio_sampling_rate {audio_sampling_rate} does not match audio sampling rate {audio.sampling_rate}"
+ )
+ audio_sampling_rate = audio.sampling_rate
+ audio = iter((audio,))
+ elif audio is not None and audio_sampling_rate is None:
+ raise ValueError("audio_sampling_rate is required when muxing an audio iterator")
+
converter = HlgGpuConverter.create(primaries, white_signal, rolloff_k, device=device)
it = iter(frames)
@@ -299,11 +307,6 @@ def encode_linear_hdr_frames_to_hlg_mp4( # noqa: PLR0913
thread_count=thread_count,
)
- audio_stream: av.audio.AudioStream | None = None
- if audio is not None:
- validate_audio_waveform(audio)
- audio_stream = prepare_audio_stream(encoder.container, audio.sampling_rate)
-
def plane_chunks() -> Iterator[tuple[np.ndarray, np.ndarray, np.ndarray]]:
y, u, v = converter(first_t)
yield y.cpu().numpy(), u.cpu().numpy(), v.cpu().numpy()
@@ -312,16 +315,26 @@ def encode_linear_hdr_frames_to_hlg_mp4( # noqa: PLR0913
yield y.cpu().numpy(), u.cpu().numpy(), v.cpu().numpy()
try:
+ audio_stream: av.audio.AudioStream | None = None
+ if audio is not None:
+ assert audio_sampling_rate is not None
+ audio_stream = prepare_audio_stream(encoder.container, audio_sampling_rate)
n = _hlg_encode_planes_threaded(
encoder,
plane_chunks(),
progress_total=progress_total if progress or progress_total is not None else None,
- audio=audio,
- audio_stream=audio_stream,
)
- except Exception:
+ mux_audio = next(audio, None) if audio is not None else None
+ if mux_audio is not None:
+ validate_audio_waveform(mux_audio)
+ assert audio_stream is not None
+ write_audio(encoder.container, audio_stream, mux_audio)
+ except BaseException:
Path(out_path).unlink(missing_ok=True)
raise
+ finally:
+ with contextlib.suppress(Exception):
+ encoder.container.close()
if progress:
logger.info(" %d frames", n)
return n
diff --git a/packages/ltx-kernels/pyproject.toml b/packages/ltx-kernels/pyproject.toml
index a510a77..f6f75e9 100644
--- a/packages/ltx-kernels/pyproject.toml
+++ b/packages/ltx-kernels/pyproject.toml
@@ -8,7 +8,7 @@ build-backend = "setuptools.build_meta"
[project]
name = "ltx-kernels"
-version = "1.4.0"
+version = "1.4.1"
dependencies = [
"torch",
"triton>=3.0",
diff --git a/packages/ltx-pipelines/pyproject.toml b/packages/ltx-pipelines/pyproject.toml
index 39f67c3..6e80567 100644
--- a/packages/ltx-pipelines/pyproject.toml
+++ b/packages/ltx-pipelines/pyproject.toml
@@ -1,6 +1,6 @@
[project]
name = "ltx-pipelines"
-version = "1.4.0"
+version = "1.4.1"
description = "Pipelines implementation for Lightricks' LTX-2 model"
readme = "README.md"
requires-python = ">=3.10"
diff --git a/packages/ltx-pipelines/src/ltx_pipelines/a2vid_two_stage.py b/packages/ltx-pipelines/src/ltx_pipelines/a2vid_two_stage.py
index 4c88181..744adb7 100644
--- a/packages/ltx-pipelines/src/ltx_pipelines/a2vid_two_stage.py
+++ b/packages/ltx-pipelines/src/ltx_pipelines/a2vid_two_stage.py
@@ -100,6 +100,7 @@ class _A2VidRunContext:
stage_1_height: int
encoded_audio_latent: torch.Tensor
source_audio: Audio
+ chunk_config: ChunkConfig
class A2VidPipelineTwoStage:
@@ -215,6 +216,7 @@ class A2VidPipelineTwoStage:
enhance_static_cache: bool,
generated_keyframes: int | Sequence[int] = 0,
decode_with_keyframes: bool = False,
+ chunk_config: ChunkConfig | None = None,
) -> _A2VidRunContext:
"""Validate inputs, decode/encode audio, and resolve frame count / tiling before chunk planning."""
images = self.image_conditioner.resolve_crf(images)
@@ -241,6 +243,9 @@ class A2VidPipelineTwoStage:
)
num_frames = snap_frames_to_grid(num_frames)
+ if chunk_config is None:
+ chunk_config = ChunkConfig(chunk_pixel_frames=num_frames, next_video_carry_frames=0)
+
ctx_p, ctx_n = self.prompt_encoder(
[prompt, negative_prompt],
enhance_first_prompt=enhance_prompt,
@@ -255,7 +260,13 @@ class A2VidPipelineTwoStage:
tiling_config,
scale_factors=scale_factors,
vae_checkpoint_path=self.video_decoder.checkpoint_path,
- video_shape=VideoPixelShape(batch=1, frames=num_frames, height=height, width=width, fps=frame_rate),
+ video_shape=VideoPixelShape(
+ batch=1,
+ frames=min(num_frames, chunk_config.chunk_pixel_frames),
+ height=height,
+ width=width,
+ fps=frame_rate,
+ ),
diffvae_optimization=self.video_decoder.diffvae_optimization,
device=self.device,
keyframes=decode_with_keyframes,
@@ -284,6 +295,7 @@ class A2VidPipelineTwoStage:
stage_1_height=height // 2,
encoded_audio_latent=encoded_audio_latent,
source_audio=source_audio,
+ chunk_config=chunk_config,
)
def _video_conditionings(
@@ -412,9 +424,8 @@ class A2VidPipelineTwoStage:
enhance_static_cache=enhance_static_cache,
generated_keyframes=generated_keyframes,
decode_with_keyframes=decode_with_keyframes,
+ chunk_config=chunk_config,
)
- if chunk_config is None:
- chunk_config = ChunkConfig(chunk_pixel_frames=ctx.num_frames, next_video_carry_frames=0)
sigmas = (
stage_1_sigmas if stage_1_sigmas is not None else self._scheduler.execute(steps=num_inference_steps)
@@ -445,7 +456,7 @@ class A2VidPipelineTwoStage:
context=VideoAudio(video=ctx.v_context_p, audio=ctx.a_context_p),
device=self.device,
dtype=ctx.dtype,
- config=chunk_config,
+ config=ctx.chunk_config,
video_scale_factors=self.stage_1.video_scale_factors,
audio_latent=ctx.encoded_audio_latent,
make_video_conditionings=self._video_conditionings(ctx.images, ctx.num_frames, color_space),
diff --git a/packages/ltx-pipelines/src/ltx_pipelines/distilled.py b/packages/ltx-pipelines/src/ltx_pipelines/distilled.py
index 3ec33f4..b80b0ee 100644
--- a/packages/ltx-pipelines/src/ltx_pipelines/distilled.py
+++ b/packages/ltx-pipelines/src/ltx_pipelines/distilled.py
@@ -153,6 +153,7 @@ class _DistilledRunContext:
tiling_config: TilingConfig
stage_1_width: int
stage_1_height: int
+ chunk_config: ChunkConfig
class DistilledPipeline:
@@ -254,6 +255,7 @@ class DistilledPipeline:
enhance_static_cache: bool,
generated_keyframes: int | Sequence[int],
decode_with_keyframes: bool,
+ chunk_config: ChunkConfig | None = None,
) -> _DistilledRunContext:
"""Validate inputs, encode prompts, and resolve frame count / tiling before chunk planning."""
require_num_frames_source(num_frames, self.duration_predictor)
@@ -285,13 +287,20 @@ class DistilledPipeline:
)
)
+ if chunk_config is None:
+ chunk_config = ChunkConfig(chunk_pixel_frames=resolved_num_frames, next_video_carry_frames=0)
+
scale_factors = tiling_scale_factors_for_vae(self.video_decoder.checkpoint_path)
resolved_tiling_config = ensure_tiling_config(
tiling_config,
scale_factors=scale_factors,
vae_checkpoint_path=self.video_decoder.checkpoint_path,
video_shape=VideoPixelShape(
- batch=1, frames=resolved_num_frames, height=height, width=width, fps=frame_rate
+ batch=1,
+ frames=min(resolved_num_frames, chunk_config.chunk_pixel_frames),
+ height=height,
+ width=width,
+ fps=frame_rate,
),
diffvae_optimization=self.video_decoder.diffvae_optimization,
device=self.device,
@@ -310,6 +319,7 @@ class DistilledPipeline:
tiling_config=resolved_tiling_config,
stage_1_width=width // 2,
stage_1_height=height // 2,
+ chunk_config=chunk_config,
)
def _sampler_kwargs(self, seed: int, noise_seed_offset: int) -> dict[str, Any]:
@@ -435,9 +445,8 @@ class DistilledPipeline:
enhance_static_cache=enhance_static_cache,
generated_keyframes=generated_keyframes,
decode_with_keyframes=decode_with_keyframes,
+ chunk_config=chunk_config,
)
- if chunk_config is None:
- chunk_config = ChunkConfig(chunk_pixel_frames=ctx.num_frames, next_video_carry_frames=0)
stage_1_sigmas = stage_1_sigmas.to(dtype=torch.float32, device=self.device)
stage_2_sigmas = stage_2_sigmas.to(dtype=torch.float32, device=self.device)
@@ -454,7 +463,7 @@ class DistilledPipeline:
context=VideoAudio(video=ctx.video_context, audio=ctx.audio_context),
device=self.device,
dtype=ctx.dtype,
- config=chunk_config,
+ config=ctx.chunk_config,
video_scale_factors=self.stage.video_scale_factors,
make_video_conditionings=self._video_conditionings(ctx.images, ctx.num_frames, color_space),
generated_keyframes=generated_keyframes,
diff --git a/packages/ltx-pipelines/src/ltx_pipelines/dubit.py b/packages/ltx-pipelines/src/ltx_pipelines/dubit.py
index b42609e..0f00d5e 100644
--- a/packages/ltx-pipelines/src/ltx_pipelines/dubit.py
+++ b/packages/ltx-pipelines/src/ltx_pipelines/dubit.py
@@ -105,6 +105,7 @@ class _DubItRunContext:
ref_audio_latent: torch.Tensor
reference_video_path: str
reference_strength: float
+ chunk_config: ChunkConfig
class DubItPipeline:
@@ -211,6 +212,7 @@ class DubItPipeline:
tiling_config: TilingConfig | AutoTiling | None,
generated_keyframes: int | Sequence[int] = 0,
decode_with_keyframes: bool = False,
+ chunk_config: ChunkConfig | None = None,
) -> _DubItRunContext:
"""Validate inputs, encode prompts/reference, and resolve frame count / tiling before chunk planning."""
images = self.image_conditioner.resolve_crf(images)
@@ -221,6 +223,9 @@ class DubItPipeline:
num_frames = snap_frames_to_grid(meta.frames)
frame_rate = float(meta.fps)
+ if chunk_config is None:
+ chunk_config = ChunkConfig(chunk_pixel_frames=num_frames, next_video_carry_frames=0)
+
generator = torch.Generator(device=self.device).manual_seed(seed)
noiser = GaussianNoiser(generator=generator)
resolved_vae_dtype = self.dtype if vae_dtype is None else vae_dtype
@@ -238,7 +243,13 @@ class DubItPipeline:
tiling_config,
scale_factors=scale_factors,
vae_checkpoint_path=self.video_decoder.checkpoint_path,
- video_shape=VideoPixelShape(batch=1, frames=num_frames, height=height, width=width, fps=frame_rate),
+ video_shape=VideoPixelShape(
+ batch=1,
+ frames=min(num_frames, chunk_config.chunk_pixel_frames),
+ height=height,
+ width=width,
+ fps=frame_rate,
+ ),
diffvae_optimization=self.video_decoder.diffvae_optimization,
device=self.device,
keyframes=decode_with_keyframes,
@@ -260,6 +271,7 @@ class DubItPipeline:
ref_audio_latent=ref_audio_latent,
reference_video_path=reference_video_path,
reference_strength=reference_strength,
+ chunk_config=chunk_config,
)
def _video_conditionings(
@@ -382,9 +394,8 @@ class DubItPipeline:
tiling_config=tiling_config,
generated_keyframes=generated_keyframes,
decode_with_keyframes=decode_with_keyframes,
+ chunk_config=chunk_config,
)
- if chunk_config is None:
- chunk_config = ChunkConfig(chunk_pixel_frames=ctx.num_frames, next_video_carry_frames=0)
stage_1_sigmas = stage_1_sigmas.to(dtype=torch.float32, device=self.device)
stage_2_sigmas = stage_2_sigmas.to(dtype=torch.float32, device=self.device)
@@ -412,7 +423,7 @@ class DubItPipeline:
context=VideoAudio(video=ctx.video_context, audio=ctx.audio_context),
device=self.device,
dtype=ctx.dtype,
- config=chunk_config,
+ config=ctx.chunk_config,
video_scale_factors=self.stage.video_scale_factors,
make_video_conditionings=make_video,
make_audio_conditionings=self._audio_conditionings(latent=ctx.ref_audio_latent, fps=ctx.frame_rate),
diff --git a/packages/ltx-pipelines/src/ltx_pipelines/ic_lora.py b/packages/ltx-pipelines/src/ltx_pipelines/ic_lora.py
index 1b89696..46baa6a 100644
--- a/packages/ltx-pipelines/src/ltx_pipelines/ic_lora.py
+++ b/packages/ltx-pipelines/src/ltx_pipelines/ic_lora.py
@@ -279,6 +279,7 @@ class _ICLoraRunContext:
conditioning_attention_mask: torch.Tensor | None
decode_with_keyframes: bool
color_space: EXRColorSpace | None
+ chunk_config: ChunkConfig
class ICLoraPipeline:
@@ -439,6 +440,7 @@ class ICLoraPipeline:
color_space: EXRColorSpace | None,
generated_keyframes: int | Sequence[int] = 0,
decode_with_keyframes: bool = False,
+ chunk_config: ChunkConfig | None = None,
) -> _ICLoraRunContext:
images = self.image_conditioner.resolve_crf(images)
assert_generated_keyframes_request(
@@ -464,6 +466,10 @@ class ICLoraPipeline:
)
resolved_num_frames = snap_frames_to_grid(num_frames)
+
+ if chunk_config is None:
+ chunk_config = ChunkConfig(chunk_pixel_frames=resolved_num_frames, next_video_carry_frames=0)
+
generator = torch.Generator(device=self.device).manual_seed(seed)
noiser = GaussianNoiser(generator=generator)
resolved_vae_dtype = self.dtype if vae_dtype is None else vae_dtype
@@ -483,7 +489,7 @@ class ICLoraPipeline:
vae_checkpoint_path=self.video_decoder.checkpoint_path,
video_shape=VideoPixelShape(
batch=1,
- frames=resolved_num_frames,
+ frames=min(resolved_num_frames, chunk_config.chunk_pixel_frames),
height=height,
width=width,
fps=frame_rate,
@@ -512,6 +518,7 @@ class ICLoraPipeline:
conditioning_attention_mask=conditioning_attention_mask,
decode_with_keyframes=decode_with_keyframes,
color_space=color_space,
+ chunk_config=chunk_config,
)
def _video_conditionings(
@@ -711,9 +718,8 @@ class ICLoraPipeline:
color_space=color_space,
generated_keyframes=generated_keyframes,
decode_with_keyframes=decode_with_keyframes,
+ chunk_config=chunk_config,
)
- if chunk_config is None:
- chunk_config = ChunkConfig(chunk_pixel_frames=ctx.num_frames, next_video_carry_frames=0)
target = VideoPixelShape(
batch=1,
@@ -737,7 +743,7 @@ class ICLoraPipeline:
context=VideoAudio(video=ctx.video_context, audio=ctx.audio_context),
device=self.device,
dtype=ctx.dtype,
- config=chunk_config,
+ config=ctx.chunk_config,
video_scale_factors=self._diffusion_stage.video_scale_factors,
generated_keyframes=generated_keyframes,
)
diff --git a/packages/ltx-pipelines/src/ltx_pipelines/ti2vid_two_stages.py b/packages/ltx-pipelines/src/ltx_pipelines/ti2vid_two_stages.py
index afe063e..8238658 100644
--- a/packages/ltx-pipelines/src/ltx_pipelines/ti2vid_two_stages.py
+++ b/packages/ltx-pipelines/src/ltx_pipelines/ti2vid_two_stages.py
@@ -100,6 +100,7 @@ class _TI2VidRunContext:
tiling_config: TilingConfig
stage_1_width: int
stage_1_height: int
+ chunk_config: ChunkConfig
class TI2VidTwoStagesPipeline:
@@ -217,6 +218,7 @@ class TI2VidTwoStagesPipeline:
enhance_static_cache: bool,
generated_keyframes: int | Sequence[int],
decode_with_keyframes: bool,
+ chunk_config: ChunkConfig | None = None,
) -> _TI2VidRunContext:
"""Validate inputs, encode prompts, and resolve frame count / tiling before chunk planning."""
require_num_frames_source(num_frames, self.duration_predictor)
@@ -248,13 +250,20 @@ class TI2VidTwoStagesPipeline:
)
)
+ if chunk_config is None:
+ chunk_config = ChunkConfig(chunk_pixel_frames=resolved_num_frames, next_video_carry_frames=0)
+
scale_factors = tiling_scale_factors_for_vae(self.video_decoder.checkpoint_path)
resolved_tiling_config = ensure_tiling_config(
tiling_config,
scale_factors=scale_factors,
vae_checkpoint_path=self.video_decoder.checkpoint_path,
video_shape=VideoPixelShape(
- batch=1, frames=resolved_num_frames, height=height, width=width, fps=frame_rate
+ batch=1,
+ frames=min(resolved_num_frames, chunk_config.chunk_pixel_frames),
+ height=height,
+ width=width,
+ fps=frame_rate,
),
diffvae_optimization=self.video_decoder.diffvae_optimization,
device=self.device,
@@ -275,6 +284,7 @@ class TI2VidTwoStagesPipeline:
tiling_config=resolved_tiling_config,
stage_1_width=width // 2,
stage_1_height=height // 2,
+ chunk_config=chunk_config,
)
def _video_conditionings(
@@ -396,9 +406,8 @@ class TI2VidTwoStagesPipeline:
enhance_static_cache=enhance_static_cache,
generated_keyframes=generated_keyframes,
decode_with_keyframes=decode_with_keyframes,
+ chunk_config=chunk_config,
)
- if chunk_config is None:
- chunk_config = ChunkConfig(chunk_pixel_frames=ctx.num_frames, next_video_carry_frames=0)
sigmas = (
stage_1_sigmas if stage_1_sigmas is not None else self._scheduler.execute(steps=num_inference_steps)
@@ -429,7 +438,7 @@ class TI2VidTwoStagesPipeline:
context=VideoAudio(video=ctx.v_context_p, audio=ctx.a_context_p),
device=self.device,
dtype=ctx.dtype,
- config=chunk_config,
+ config=ctx.chunk_config,
video_scale_factors=self.stage_1.video_scale_factors,
make_video_conditionings=self._video_conditionings(ctx.images, ctx.num_frames, color_space),
generated_keyframes=generated_keyframes,
diff --git a/packages/ltx-pipelines/src/ltx_pipelines/utils/media_io/encode.py b/packages/ltx-pipelines/src/ltx_pipelines/utils/media_io/encode.py
index ae07fc4..bd70b14 100644
--- a/packages/ltx-pipelines/src/ltx_pipelines/utils/media_io/encode.py
+++ b/packages/ltx-pipelines/src/ltx_pipelines/utils/media_io/encode.py
@@ -30,56 +30,55 @@ logger = logging.getLogger(__name__)
def _encode_hdr_video_outputs(
- chunks: Iterator[torch.Tensor],
+ video: Iterator[torch.Tensor],
output_path: Path,
fps: float,
color_space: EXRColorSpace,
*,
max_workers: int = 8,
audio: Iterator[Audio] | None = None,
+ audio_sampling_rate: int | None = None,
device: torch.device | None = None,
) -> Path:
"""Write EXR frames plus a BT.2020/HLG master.
- One pass over ``chunks`` feeds two independent sinks:
+ One pass over ``video`` feeds two independent sinks:
* **HLG** — always Rec.709 scene-linear (``decode_hdr_video`` default).
* **EXR** — matches ``color_space``: log codes when :attr:`~EXRColorSpace.is_log_working`,
else scene-linear in ``color_space.source_primaries`` (reuse the HLG tensor when that
is already Rec.709).
- Chunks are collected first so a stitch iterator can be pulled before HLG
- (which takes a single :class:`~ltx_core.types.Audio`). EXR and HLG then share
- that list.
+ Video tensors flow directly from decode through both sinks. When ``audio`` is
+ deferred, HLG opens its AAC stream from ``audio_sampling_rate`` and pulls
+ the completed track only after this iterator is exhausted.
Args:
- chunks: Per-chunk VAE decode output, each ``[F, H, W, C]`` float in ``[0, 1]``
- (the compressed working-space signal).
+ video: VAE decode output as an iterator of ``[F, H, W, C]`` float tensors in
+ ``[0, 1]`` (the compressed working-space signal), e.g. from
+ :func:`~ltx_pipelines.chunks.split_decoded_chunks`.
output_path: Destination of the HLG master. EXR frames go to
``<stem>_<label>_exr/`` where ``label`` is :func:`~ltx_pipelines.utils.media_io.exr.exr_dir_label`.
fps: Output frame rate.
color_space: HDR colour space for EXR tags / linear EXR primaries.
max_workers: Number of EXR writer threads.
- audio: Optional stereo track muxed into the HLG master. Drain of ``chunks``
- fills a lazy stitch iterator; pull it before HLG.
+ audio: Optional stereo track muxed into the HLG master. It may be filled
+ lazily while ``video`` is consumed.
+ audio_sampling_rate: AAC sample rate, required when ``audio`` is set.
device: Device for the HLG colour conversion. Defaults to CUDA when available.
Returns:
The directory the EXR frames were written to.
Raises:
- ValueError: If ``chunks`` yields nothing.
+ ValueError: If ``video`` yields nothing.
"""
exr_dir = output_path.parent / f"{output_path.stem}_{exr_dir_label(color_space)}_exr"
exr_dir.mkdir(parents=True, exist_ok=True)
exr_primaries = color_space.source_primaries
exr_color_space = exr_colorspace_tag(color_space)
frames_written = 0
- chunk_list = list(chunks)
- mux_audio = next(audio, None) if audio is not None else None
- if mux_audio is not None:
- validate_audio_waveform(mux_audio)
def hlg_chunks() -> Iterator[torch.Tensor]:
"""Write EXR for each chunk, then yield Rec.709 linear for the HLG encoder."""
nonlocal frames_written
with ThreadPoolExecutor(max_workers=max_workers) as executor:
pending: deque[Future[None]] = deque()
- for chunk in chunk_list:
+ for chunk in video:
# HLG sink: always Rec.709 scene-linear.
hlg_linear = decode_hdr_video(chunk)
@@ -108,17 +107,18 @@ def _encode_hdr_video_outputs(
logger.info(
"Encoding BT.2020/HLG/10-bit master%s: %s",
- " (with audio)" if mux_audio is not None else "",
+ " (with audio)" if audio is not None else "",
output_path,
)
- # Empty ``chunks`` raises inside the HLG encoder (after EXR side effects are none).
+ # Empty ``video`` raises inside the HLG encoder (after EXR side effects are none).
encode_linear_hdr_frames_to_hlg_mp4(
hlg_chunks(),
output_path,
fps=fps,
primaries=Primaries.REC709,
- audio=mux_audio,
+ audio=audio,
device=device,
+ audio_sampling_rate=audio_sampling_rate,
)
logger.info(
"Wrote %d EXR frames (%s/%s) to %s",
@@ -248,9 +248,8 @@ def encode_video( # noqa: PLR0913
or None for a video-only file. Pass
:func:`~ltx_pipelines.chunks.deferred_stitch_audio` when parts are collected
while ``video`` is consumed, and set ``audio_sampling_rate`` so AAC can open
- before any video packets (SDR). HDR drains ``video`` first, then pulls that
- iterator and passes one :class:`~ltx_core.types.Audio` to HLG. Waveforms must
- be stereo ``(2, N)`` or ``(N, 2)``.
+ before any video packets. Both SDR and HDR pull the iterator only after
+ draining ``video``. Waveforms must be stereo ``(2, N)`` or ``(N, 2)``.
output_path: Destination path. Parent directories are created if missing.
Partial SDR output is removed if encoding fails. HDR writes EXR under
``<stem>_<label>_exr/`` (see :func:`~ltx_pipelines.utils.media_io.exr.exr_dir_label`) and an
@@ -263,7 +262,7 @@ def encode_video( # noqa: PLR0913
color_space: HDR colour space for the encode sink. ``None`` (default) is SDR.
When set, writes half EXR frames + HLG master.
audio_sampling_rate: AAC sample rate. Inferred when ``audio`` is an
- :class:`~ltx_core.types.Audio`; required for an audio iterator on the SDR path.
+ :class:`~ltx_core.types.Audio`; required for an audio iterator.
Returns:
EXR output directory when HDR, else ``None``.
Raises:
@@ -271,6 +270,7 @@ def encode_video( # noqa: PLR0913
explicit and embedded audio sampling rates.
"""
if isinstance(audio, Audio):
+ validate_audio_waveform(audio)
if audio_sampling_rate is not None and audio_sampling_rate != audio.sampling_rate:
raise ValueError(
f"audio_sampling_rate {audio_sampling_rate} does not match audio sampling rate {audio.sampling_rate}"
@@ -288,6 +288,7 @@ def encode_video( # noqa: PLR0913
fps=float(fps),
color_space=color_space,
audio=audio,
+ audio_sampling_rate=audio_sampling_rate,
)
encode_sdr_h264(
diff --git a/packages/ltx-trainer/pyproject.toml b/packages/ltx-trainer/pyproject.toml
index 8294196..6a79f61 100644
--- a/packages/ltx-trainer/pyproject.toml
+++ b/packages/ltx-trainer/pyproject.toml
@@ -1,6 +1,6 @@
[project]
name = "ltx-trainer"
-version = "1.4.0"
+version = "1.4.1"
description = "LTX-2 training, democratized."
readme = "README.md"
authors = [