[GenAI] Fix CI and stability for LTX-Video LoRA support (Supersedes #3309)

[goyaladitya05]

This PR builds upon the work started by @Eshaan-byte in #3309. I've integrated his logic and added the fixes for the set_adapters bindings and AdapterController initialization to resolve the CI issues.

Changes:

• Added binding for LTXVideoTransformer3DModel.set_adapters() to resolve AttributeError in the TestLoRAVideoGeneration test suite.
• Added a safety check in set_adapters() to prevent segmentation faults when the AdapterController is not initialized (e.g., when no adapters are provided at construction).

Closes #3306
Supersedes #3309

Verification Results:

without LoRA

with LoRA (Adapter: ltxv_095_cakeify_lora.safetensors)

These samples demonstrate that the LoRA adapter is successfully loaded and applied during the denoising process in the Text2VideoPipeline.

Environment:
Model: LTX-Video
Alpha: 1.5
Prompt:"A cute golden retriever puppy running in a green grassy field on a sunny day, high quality, photorealistic"

Checklist:

• This PR follows GenAI Contributing guidelines.
• Tests have been updated or added to cover the new code.
  • Verified tests/python_tests/test_video_generation.py passes locally.
• This PR fully addresses the ticket.
• I have made corresponding changes to the documentation.

[Copilot]

Pull request overview

This PR fixes CI failures from #3309 by adding the missing Python binding for LTXVideoTransformer3DModel.set_adapters() and implementing a safety check to prevent segmentation faults when the AdapterController is not initialized. The changes enable LoRA adapter support for LTX-Video generation, following the same architectural pattern used in image generation models.

Changes:

• Added Python binding for set_adapters() method on the LTXVideoTransformer3DModel
• Implemented safety check in C++ to prevent crashes when calling set_adapters() on models compiled without adapter support
• Added comprehensive tests for LoRA functionality in the video generation pipeline

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/cpp/src/video_generation/models/ltx_video_transformer_3d_model.cpp	Removed blocking assertion for adapters, added adapter initialization in compile(), implemented set_adapters() with safety check
src/cpp/include/openvino/genai/video_generation/ltx_video_transformer_3d_model.hpp	Added lora_adapter.hpp include, declared set_adapters() method, added AdapterController member
src/python/py_video_generation_models.cpp	Added Python binding for set_adapters() method with proper documentation
tests/python_tests/test_video_generation.py	Added TestLoRAVideoGeneration class with tests for adapter functionality in constructor, generate(), and set_adapters()
samples/cpp/video_generation/lora_text2video.cpp	New sample demonstrating LoRA usage with video generation, following the pattern from lora_text2image.cpp
samples/cpp/video_generation/CMakeLists.txt	Added build configuration for lora_text2video sample, consistent with image generation samples

Comments suppressed due to low confidence (1)

tests/python_tests/test_video_generation.py:270

• The test test_generate_without_cfg_after_reshape_with_cfg (lines 256-270) doesn't use any LoRA adapters functionality, so it seems misplaced in the TestLoRAVideoGeneration class. This test appears to be testing CFG (Classifier-Free Guidance) behavior, not LoRA functionality. Consider moving it to a more appropriate test class (e.g., TestText2VideoPipelineAdvanced) or creating a separate TestCFGBehavior class for CFG-related tests.

    def test_generate_without_cfg_after_reshape_with_cfg(self, video_generation_model):
        """Test: reshape with CFG then generate without CFG should raise error."""
        pipe = ov_genai.Text2VideoPipeline(video_generation_model)
        pipe.reshape(1, 9, 32, 32, 3.0)
        pipe.compile("CPU")

        with pytest.raises(RuntimeError, match="guidance_scale <= 1 requested, but the compiled model expects CFG"):
            pipe.generate(
                "test prompt",
                guidance_scale=1.0,
                height=32,
                width=32,
                num_frames=9,
                num_inference_steps=2,
            )

[goyaladitya05]

Thanks for the feedback.
I’ve addressed the review comments (moved the CFG test, refined the ASSERT message) and rebased on master.
Could someone please run the CI once?

[likholat]

Could you please also attach two short videos (converted to GIF) showing samples generated with and without LoRA adapter? Please add them to the PR description

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

[goyaladitya05]

Thanks for the feedback. I've updated the PR to address the review comments.

• src/cpp/src/video_generation/ltx_pipeline.hpp:
  Added the missing m_transformer->set_adapters() invocation.

• src/cpp/include/openvino/genai/video_generation/generation_config.hpp & .../generation_config_utils.cpp:
  Added the adapters field to VideoGenerationConfig to mirror the image generation pipelines.

• src/python/py_video_generation_models.cpp: Exposed set_adapters to the Python bindings.

Samples & Documentation

• samples/python/video_generation/lora_text2video.py: Created the missing Python sample.
• samples/cpp/video_generation/lora_text2video.cpp: Cleaned the C++ logic.
• samples/cpp/video_generation/CMakeLists.txt: Connected the new C++ sample.
• samples/.../video_generation/README.md (both C++ & Python): Added run commands and documentation for the new LoRA samples, keeping the original example prompts used earlier.

Unit Tests

• tests/python_tests/test_video_generation.py:
  Restored test_generate_without_cfg_default_compile which was accidentally lost in a rebase, and moved the two CFG to the correct location.

[goyaladitya05]

Could you please also attach two short videos (converted to GIF) showing samples generated with and without LoRA adapter? Please add them to the PR description

@likholat , I've added the samples to the PR description. Please review whenever you have the time.