NVIDIA · shy982 · Jun 22, 2026
diff --git a/README.md b/README.md
@@ -99,6 +99,7 @@ runner slugs, multi-GPU commands, and (where available) profiling benchmarks.
 | [Causal Wan2.2](https://nvidia.github.io/flashdreams/main/models/causal_wan22.html) | FastVideo Causal Wan 2.2 14B MoE T2V |
 | [FlashVSR](https://nvidia.github.io/flashdreams/main/models/flashvsr.html) | Streaming video super-resolution |
 | [Cosmos-Predict2.5](https://nvidia.github.io/flashdreams/main/models/cosmos_predict2.html) | Bidirectional T2V / I2V |
+| [LTX-Video](https://nvidia.github.io/flashdreams/main/models/ltx_video.html) | Causal streaming T2V (2B) |
 
 See [the model gallery](https://nvidia.github.io/flashdreams/main/models/index.html) and
 [the new method guide](https://nvidia.github.io/flashdreams/main/developer_guides/new_integration.html)

diff --git a/docs/source/api/integrations.rst b/docs/source/api/integrations.rst
@@ -58,6 +58,7 @@ Reference integration folders
 - `fastvideo_causal_wan22 <https://github.com/NVIDIA/flashdreams/tree/main/integrations/fastvideo_causal_wan22>`_
 - `flashvsr <https://github.com/NVIDIA/flashdreams/tree/main/integrations/flashvsr>`_
 - `cosmos_predict2 <https://github.com/NVIDIA/flashdreams/tree/main/integrations/cosmos_predict2>`_
+- `ltx_video <https://github.com/NVIDIA/flashdreams/tree/main/integrations/ltx_video>`_
 
 NVIDIA OmniDreams
 -----------------

diff --git a/docs/source/index.rst b/docs/source/index.rst
@@ -232,6 +232,13 @@ invocation, the checkpoint source, and the per-implementation knobs.
 
       Autoregressive text-to-video based on Wan 2.2 from FastVideo.
 
+   .. grid-item-card:: LTX-Video
+      :class-card: fd-feature
+      :link: models/ltx_video
+      :link-type: doc
+
+      Causal 2B streaming text-to-video with optional KV-cache and compile path.
+
    .. grid-item-card:: LingBot-World
       :class-card: fd-feature
       :link: models/lingbot_world

diff --git a/docs/source/models/index.rst b/docs/source/models/index.rst
@@ -27,6 +27,7 @@ Models
    cosmos_predict2
    flashvsr
    lingbot_world
+   ltx_video
    wan21
 
 FlashDreams runs a growing family of world and video models (text-to-video,
@@ -106,6 +107,13 @@ uses, and the settings you can tune.
 
       Autoregressive text-to-video based on Wan 2.2 from FastVideo.
 
+   .. grid-item-card:: LTX-Video
+      :class-card: fd-feature
+      :link: /models/ltx_video
+      :link-type: doc
+
+      Causal 2B streaming text-to-video with optional KV-cache and compile path.
+
    .. grid-item-card:: LingBot-World
       :class-card: fd-feature
       :link: /models/lingbot_world

diff --git a/docs/source/models/ltx_video.rst b/docs/source/models/ltx_video.rst
@@ -0,0 +1,92 @@
+.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+.. SPDX-License-Identifier: Apache-2.0
+..
+.. Licensed under the Apache License, Version 2.0 (the "License");
+.. you may not use this file except in compliance with the License.
+.. You may obtain a copy of the License at
+..
+.. http://www.apache.org/licenses/LICENSE-2.0
+..
+.. Unless required by applicable law or agreed to in writing, software
+.. distributed under the License is distributed on an "AS IS" BASIS,
+.. WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+.. See the License for the specific language governing permissions and
+.. limitations under the License.
+
+LTX-Video
+===================================
+
+.. raw:: html
+
+   <div class="model-link-row">
+     <a class="model-link-button" href="https://github.com/Lightricks/LTX-Video" target="_blank" rel="noopener noreferrer">Official code</a>
+     <a class="model-link-button" href="https://huggingface.co/Lightricks/LTX-Video" target="_blank" rel="noopener noreferrer">Model weights</a>
+   </div>
+
+`LTX-Video <https://github.com/Lightricks/LTX-Video>`_ is a 2B causal video diffusion
+model with a causal VAE suited to autoregressive chunk generation. This integration
+wraps ``LTXPipeline`` from ``diffusers`` and adds an optional optimized path with
+manual denoising, KV-cache, ``torch.compile``, and FlashAttention.
+
+Requirements
+------------
+
+- **Minimum VRAM**: ~24 GB for the streaming runner; ~40 GB+ for the optimized
+  stack at 768×512 with KV-cache enabled.
+- **PyTorch**: >= 2.9 (CUDA 13.x recommended; see :doc:`/quickstart/installation`).
+- **diffusers**: recent release with ``LTXPipeline`` support.
+
+Installation
+------------
+
+.. code-block:: bash
+
+   # from the repo root
+   uv sync --project integrations/ltx_video
+
+Running the method
+------------------
+
+Launch one of the registered runner slugs via ``flashdreams-run``:
+
+.. code-block:: bash
+
+   export HF_TOKEN=<your-hf-token>
+
+   uv run --project integrations/ltx_video \
+       flashdreams-run \
+       ltx-video-t2v-2b \
+       --prompt "A coastal road at dusk, waves breaking on rocky cliffs, cinematic wide shot" \
+       --pixel-height 512 --pixel-width 768 \
+       --total-blocks 7
+
+We provide the following variants:
+
+.. list-table::
+   :header-rows: 1
+   :widths: 45 55
+
+   * - Method
+     - Description
+   * - ``ltx-video-t2v-2b``
+     - Streaming ``LTXPipeline`` wrapper — fast startup, native ``pipe()`` per chunk.
+   * - ``ltx-video-t2v-2b-optimized``
+     - Manual denoise loop with KV-cache, ``torch.compile``, and FlashAttention.
+   * - ``ltx-video-t2v-2b-taehv``
+     - Optimized path plus TAEHV fast decoder.
+
+The optimized runner improves time-to-first-frame after warmup; steady-state
+throughput tuning is ongoing.
+
+Tests
+-----
+
+.. code-block:: bash
+
+   uv run --project integrations/ltx_video pytest integrations/ltx_video/tests/test_smoke.py -v
+
+GPU optimization tests (requires CUDA + weights):
+
+.. code-block:: bash
+
+   uv run --project integrations/ltx_video pytest integrations/ltx_video/tests/test_optimizations.py -v -m ci_gpu
diff --git a/integrations/ltx_video/README.md b/integrations/ltx_video/README.md
@@ -0,0 +1,46 @@
+# LTX-Video × FlashDreams
+
+First-party FlashDreams integration for [LTX-Video](https://github.com/Lightricks/LTX-Video) 2B causal streaming text-to-video.
+
+## Runner slugs
+
+| Slug | Mode |
+|------|------|
+| `ltx-video-t2v-2b` | Streaming `LTXPipeline` wrapper (native `pipe()` per chunk) |
+| `ltx-video-t2v-2b-optimized` | Manual denoise + KV-cache + `torch.compile` + FlashAttention |
+| `ltx-video-t2v-2b-taehv` | Optimized path + TAEHV fast decoder |
+
+## Install
+
+```bash
+# from the repo root
+export HF_TOKEN=<your-hf-token>
+uv sync --project integrations/ltx_video
+```
+
+## Run
+
+```bash
+uv run --project integrations/ltx_video \
+    flashdreams-run ltx-video-t2v-2b \
+    --prompt "A coastal road at dusk, waves breaking on rocky cliffs, cinematic wide shot" \
+    --pixel-height 512 --pixel-width 768 \
+    --total-blocks 7
+```
+
+Optimized path:
+
+```bash
+uv run --project integrations/ltx_video \
+    flashdreams-run ltx-video-t2v-2b-optimized \
+    --total-blocks 7
+```
+
+## Tests
+
+```bash
+uv run --project integrations/ltx_video pytest integrations/ltx_video/tests/test_smoke.py -v
+uv run --project integrations/ltx_video pytest integrations/ltx_video/tests/test_optimizations.py -v -m ci_gpu
+```
+
+See the [LTX-Video model page](https://nvidia.github.io/flashdreams/main/models/ltx_video.html) for full setup notes.
diff --git a/integrations/ltx_video/config/ltx-video-t2v-2b-optimized.yaml b/integrations/ltx_video/config/ltx-video-t2v-2b-optimized.yaml
@@ -0,0 +1,20 @@
+# LTX-Video 2B T2V — full optimization stack
+runner: ltx-video-t2v-2b-optimized
+pipeline:
+  _target_: ltx_video.pipeline.LTXVideoStreamingPipeline
+checkpoint: Lightricks/LTX-Video
+generation:
+  chunk_frames: 25
+  chunk_overlap: 1
+  total_blocks: 7
+  num_inference_steps: 50
+  width: 768
+  height: 512
+  fps: 24
+  kv_cache: true
+  kv_window_size: null
+  compile: true
+  cuda_graphs: true
+  manual_denoise: true
+  flash_attention: true
+  use_taehv: false
diff --git a/integrations/ltx_video/config/ltx-video-t2v-2b-taehv.yaml b/integrations/ltx_video/config/ltx-video-t2v-2b-taehv.yaml
@@ -0,0 +1,18 @@
+# Optimized + TAEHV fast decoder variant
+runner: ltx-video-t2v-2b-taehv
+pipeline:
+  _target_: ltx_video.pipeline.LTXVideoStreamingPipeline
+checkpoint: Lightricks/LTX-Video
+generation:
+  chunk_frames: 25
+  chunk_overlap: 1
+  total_blocks: 7
+  num_inference_steps: 50
+  width: 768
+  height: 512
+  fps: 24
+  kv_cache: true
+  compile: true
+  cuda_graphs: true
+  flash_attention: true
+  use_taehv: true
diff --git a/integrations/ltx_video/config/ltx-video-t2v-2b.yaml b/integrations/ltx_video/config/ltx-video-t2v-2b.yaml
@@ -0,0 +1,17 @@
+# LTX-Video 2B T2V — streaming only (native pipe() per chunk)
+runner: ltx-video-t2v-2b
+pipeline:
+  _target_: ltx_video.pipeline.LTXVideoStreamingPipeline
+checkpoint: Lightricks/LTX-Video
+generation:
+  chunk_frames: 25
+  chunk_overlap: 1
+  total_blocks: 7
+  num_inference_steps: 50
+  width: 768
+  height: 512
+  fps: 24
+  kv_cache: false
+  compile: false
+  cuda_graphs: false
+  flash_attention: true
diff --git a/integrations/ltx_video/ltx_video/__init__.py b/integrations/ltx_video/ltx_video/__init__.py
@@ -0,0 +1,8 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""FlashDreams LTX-Video full optimization integration."""
+
+from ltx_video.pipeline import LTXVideoStreamingPipeline
+
+__all__ = ["LTXVideoStreamingPipeline"]