Video-to-Video (V2V)

Video-to-Video (V2V) is a workflow in which an existing video provides visual information that guides the generation of a new video. The official LTX-2 training repository explicitly provides an IC-LoRA (V2V) training mode for video-to-video transformations. Its implementation uses reference video conditioning: the reference video's encoded latents are supplied alongside the target sequence, allowing the model to learn transformations from the reference video toward the target output. The repository describes this approach as useful for transformations such as depth-to-video, pose control, and style transfer.

The official LTX-2 implementation describes V2V IC-LoRA training as reference-video conditioning. Reference video latents are combined with the target sequence, while the reference portion remains clean and the generated target portion is the part optimized during training. The training configuration also defines the reference video as a conditioning input that guides the transformation. This architecture allows an IC-LoRA adapter to learn how information from a reference video should influence the generated result.

The official LTX-2 V2V training configuration gives several examples of transformation-oriented use cases, including depth-to-video, pose control, and style transfer. The broader LTX documentation also describes IC-LoRA control models as a way to provide precise control over video generation. This means V2V is best understood as a controlled transformation framework rather than a single fixed visual effect. Different IC-LoRA adapters can be trained for different types of reference information and transformations.

IC-LoRA stands for In-Context LoRA. In the LTX-2 training repository, there is a dedicated v2v_ic_lora.yaml configuration for training IC-LoRA adapters for video-to-video transformations. The configuration uses reference video latents as conditioning and trains the model to transform the reference according to the target data and text prompt. This makes IC-LoRA the documented mechanism behind the V2V training workflow in the open LTX-2 ecosystem.

In Image-to-Video (I2V), an image provides the visual starting point for video generation. In the IC-LoRA V2V approach, an existing video provides the reference information used to guide a transformation. These workflows therefore solve different problems: I2V is primarily about animating or generating motion from an image, while V2V is intended for transformations driven by video-based conditioning. When the desired transformation depends on information contained across multiple frames, video conditioning provides a different form of control from a single starting image.

The open-source LTX-2 repository includes V2V-specific training configurations and inference pipelines associated with IC-LoRA. The official LTX pipeline documentation also includes video-to-video pipelines, including an HDR IC-LoRA pipeline for video-to-video generation. At the same time, the LTX API documentation does not present generic V2V as a standalone API operation; its documented video operations include text-to-video, image-to-video, audio-to-video, retake, extend, and reframe. Therefore, V2V should be treated primarily as an open-source and IC-LoRA-based workflow rather than as a separate standard API modality.

V2V is most relevant when the source video itself is an important part of the conditioning signal. This includes workflows based on visual properties such as pose, depth, or a learned style transformation. If the task only requires animating a single still image, I2V is the more directly documented modality. If the task requires changing an existing video through a learned reference-video transformation, an IC-LoRA V2V workflow is the more appropriate part of the LTX ecosystem to investigate.

The documented LTX implementation does not define one universal V2V transformation that can perform every possible video edit. Instead, IC-LoRA provides a framework for learning specific transformations from reference-video conditioning. As a result, the available behavior depends on the particular adapter, conditioning information, training setup, and inference pipeline. This is an important distinction when planning an LTX workflow: a V2V-capable pipeline does not automatically imply arbitrary video editing, masking, object replacement, or unrestricted style transfer.