feat(diffusers): add support to LTX-2

backend/python/diffusers/backend.py

@@ -41,6 +41,14 @@
 from transformers import T5EncoderModel
 from safetensors.torch import load_file
 
+# Import LTX-2 specific utilities
+try:
+    from diffusers.pipelines.ltx2.export_utils import encode_video as ltx2_encode_video
+    LTX2_AVAILABLE = True
+except ImportError:
+    LTX2_AVAILABLE = False
+    ltx2_encode_video = None
+
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
 COMPEL = os.environ.get("COMPEL", "0") == "1"
 XPU = os.environ.get("XPU", "0") == "1"
@@ -290,6 +298,20 @@ def _load_pipeline(self, request, modelFile, fromSingleFile, torchType, variant)
                 pipe.enable_model_cpu_offload()
             return pipe
 
+        # LTX2ImageToVideoPipeline - needs img2vid flag, CPU offload, and special handling
+        if pipeline_type == "LTX2ImageToVideoPipeline":
+            self.img2vid = True
+            self.ltx2_pipeline = True
+            pipe = load_diffusers_pipeline(
+                class_name="LTX2ImageToVideoPipeline",
+                model_id=request.Model,
+                torch_dtype=torchType,
+                variant=variant
+            )
+            if not DISABLE_CPU_OFFLOAD:
+                pipe.enable_model_cpu_offload()
+            return pipe
+
         # ================================================================
         # Dynamic pipeline loading - the default path for most pipelines
         # Uses the dynamic loader to instantiate any pipeline by class name
@@ -404,6 +426,7 @@ def LoadModel(self, request, context):
             fromSingleFile = request.Model.startswith("http") or request.Model.startswith("/") or local
             self.img2vid = False
             self.txt2vid = False
+            self.ltx2_pipeline = False
 
             # Load pipeline using dynamic loader
             # Special cases that require custom initialization are handled first
@@ -686,7 +709,44 @@ def GenerateVideo(self, request, context):
             print(f"Generating video with {kwargs=}", file=sys.stderr)
 
             # Generate video frames based on pipeline type
-            if self.PipelineType == "WanPipeline":
+            if self.ltx2_pipeline or self.PipelineType == "LTX2ImageToVideoPipeline":
+                # LTX-2 image-to-video generation with audio
+                if not LTX2_AVAILABLE:
+                    return backend_pb2.Result(success=False, message="LTX-2 pipeline requires diffusers.pipelines.ltx2.export_utils")
+
+                # LTX-2 uses 'image' parameter instead of 'start_image'
+                if request.start_image:
+                    image = load_image(request.start_image)
+                    kwargs["image"] = image
+                    # Remove start_image if it was added
+                    kwargs.pop("start_image", None)
+
+                # LTX-2 uses 'frame_rate' instead of 'fps'
+                frame_rate = float(fps)
+                kwargs["frame_rate"] = frame_rate
+
+                # LTX-2 requires output_type="np" and return_dict=False
+                kwargs["output_type"] = "np"
+                kwargs["return_dict"] = False
+
+                # Generate video and audio
+                video, audio = self.pipe(**kwargs)
+
+                # Convert video to uint8 format
+                video = (video * 255).round().astype("uint8")
+                video = torch.from_numpy(video)
+
+                # Use LTX-2's encode_video function which handles audio
+                ltx2_encode_video(
+                    video[0],
+                    fps=frame_rate,
+                    audio=audio[0].float().cpu(),
+                    audio_sample_rate=self.pipe.vocoder.config.output_sampling_rate,
+                    output_path=request.dst,
+                )
+
+                return backend_pb2.Result(message="Video generated successfully", success=True)
+            elif self.PipelineType == "WanPipeline":
                 # WAN2.2 text-to-video generation
                 output = self.pipe(**kwargs)
                 frames = output.frames[0]  # WAN2.2 returns frames in this format
@@ -727,7 +787,7 @@ def GenerateVideo(self, request, context):
             else:
                 return backend_pb2.Result(success=False, message=f"Pipeline {self.PipelineType} does not support video generation")
 
-            # Export video
+            # Export video (for non-LTX-2 pipelines)
             export_to_video(frames, request.dst, fps=fps)
 
             return backend_pb2.Result(message="Video generated successfully", success=True)

gallery/index.yaml

@@ -1237,6 +1237,63 @@
       cuda: true
       pipeline_type: QwenImageEditPipeline
       enable_parameters: num_inference_steps,image
+- &ltx2
+  name: "ltx-2"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/Lightricks/LTX-2
+  license: ltx-2-community-license-agreement
+  tags:
+    - diffusers
+    - gpu
+    - image-to-video
+    - video-generation
+    - audio-video
+  description: |
+    **LTX-2** is a DiT-based audio-video foundation model designed to generate synchronized video and audio within a single model. It brings together the core building blocks of modern video generation, with open weights and a focus on practical, local execution.
+
+    **Key Features:**
+    - **Joint Audio-Video Generation**: Generates synchronized video and audio in a single model
+    - **Image-to-Video**: Converts static images into dynamic videos with matching audio
+    - **High Quality**: Produces realistic video with natural motion and synchronized audio
+    - **Open Weights**: Available under the LTX-2 Community License Agreement
+
+    **Model Details:**
+    - **Model Type**: Diffusion-based audio-video foundation model
+    - **Architecture**: DiT (Diffusion Transformer) based
+    - **Developed by**: Lightricks
+    - **Paper**: [LTX-2: Efficient Joint Audio-Visual Foundation Model](https://arxiv.org/abs/2601.03233)
+
+    **Usage Tips:**
+    - Width & height settings must be divisible by 32
+    - Frame count must be divisible by 8 + 1 (e.g., 9, 17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97, 105, 113, 121)
+    - Recommended settings: width=768, height=512, num_frames=121, frame_rate=24.0
+    - For best results, use detailed prompts describing motion and scene dynamics
+
+    **Limitations:**
+    - This model is not intended or able to provide factual information
+    - Prompt following is heavily influenced by the prompting-style
+    - When generating audio without speech, the audio may be of lower quality
+
+    **Citation:**
+    ```bibtex
+    @article{hacohen2025ltx2,
+      title={LTX-2: Efficient Joint Audio-Visual Foundation Model},
+      author={HaCohen, Yoav and Brazowski, Benny and Chiprut, Nisan and others},
+      journal={arXiv preprint arXiv:2601.03233},
+      year={2025}
+    }
+    ```
+  overrides:
+    backend: diffusers
+    low_vram: true
+    parameters:
+      model: Lightricks/LTX-2
+    diffusers:
+      cuda: true
+      pipeline_type: LTX2ImageToVideoPipeline
+    options:
+      - torch_dtype:bf16
 - &gptoss
   name: "gpt-oss-20b"
   url: "github:mudler/LocalAI/gallery/harmony.yaml@master"
@@ -3774,7 +3831,7 @@
     - gemma3
     - gemma-3
   overrides:
     #mmproj: gemma-3-27b-it-mmproj-f16.gguf
     parameters:
       model: gemma-3-27b-it-Q4_K_M.gguf
   files:
@@ -3792,7 +3849,7 @@
   description: |
     google/gemma-3-12b-it is an open-source, state-of-the-art, lightweight, multimodal model built from the same research and technology used to create the Gemini models. It is capable of handling text and image input and generating text output. It has a large context window of 128K tokens and supports over 140 languages. The 12B variant has been fine-tuned using the instruction-tuning approach. Gemma 3 models are suitable for a variety of text generation and image understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes them deployable in environments with limited resources such as laptops, desktops, or your own cloud infrastructure.
   overrides:
     #mmproj: gemma-3-12b-it-mmproj-f16.gguf
     parameters:
       model: gemma-3-12b-it-Q4_K_M.gguf
   files:
@@ -3810,7 +3867,7 @@
   description: |
     Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma 3 models are multimodal, handling text and image input and generating text output, with open weights for both pre-trained variants and instruction-tuned variants. Gemma 3 has a large, 128K context window, multilingual support in over 140 languages, and is available in more sizes than previous versions. Gemma 3 models are well-suited for a variety of text generation and image understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone. Gemma-3-4b-it is a 4 billion parameter model.
   overrides:
     #mmproj: gemma-3-4b-it-mmproj-f16.gguf
     parameters:
       model: gemma-3-4b-it-Q4_K_M.gguf
   files:
@@ -7028,7 +7085,7 @@
       sha256: 2756551de7d8ff7093c2c5eec1cd00f1868bc128433af53f5a8d434091d4eb5a
       uri: huggingface://Triangle104/Nano_Imp_1B-Q8_0-GGUF/nano_imp_1b-q8_0.gguf
 - &smollm
   url: "github:mudler/LocalAI/gallery/chatml.yaml@master" ## SmolLM
   name: "smollm-1.7b-instruct"
   icon: https://huggingface.co/datasets/HuggingFaceTB/images/resolve/main/banner_smol.png
   tags:
@@ -7072,7 +7129,7 @@
       sha256: decd2598bc2c8ed08c19adc3c8fdd461ee19ed5708679d1c54ef54a5a30d4f33
       uri: huggingface://HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF/smollm2-1.7b-instruct-q4_k_m.gguf
 - &llama31
   url: "github:mudler/LocalAI/gallery/llama3.1-instruct.yaml@master" ## LLama3.1
   icon: https://avatars.githubusercontent.com/u/153379578
   name: "meta-llama-3.1-8b-instruct"
   license: llama3.1