MiniMax H3 CrossView-Warp LoRA: Change Camera Movement and Perspective in Existing Videos

An Open-Source Solution for "Re-Shooting" Existing Videos

A persistent pain point in video generation has long been this: once a clip is generated, its camera movement and perspective are essentially locked in. Changing the angle or adjusting pan, tilt, or zoom typically means regenerating the whole thing — which wastes compute and makes it nearly impossible to maintain visual consistency. MiniMax-H3 CrossView-Warp LoRA, recently released by developer Cseti, is a tool built specifically for this scenario. It lets creators directly modify the camera movement and perspective of existing video clips.

According to the developer's post on Reddit, this is the H3 version of their CrossView-Warp LoRA series (previous versions were based on the LTX model). The LoRA is a Video-to-Video (V2V) type tool, with the core capability of recomputing different camera paths while preserving the original subject content of the footage.

Components and How to Use It

The solution consists of three components, all open-sourced on Hugging Face and GitHub:

The LoRA Model

The LoRA weights and training details are hosted on Hugging Face: Cseti/MiniMax-H3_Ref2VA-LoRA-CrossView-Warp_v1. Fine-tuned on the MiniMax H3 video model, the naming convention (Ref2VA) suggests it uses a reference-frame-guided approach to perform cross-view transformation.

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique. Its core idea is to attach a pair of low-rank matrices alongside the weight matrices of a pre-trained model, training only this small pair while freezing the original weights. The benefits are significant: the number of trainable parameters is tiny (typically less than 1% of the original model), dramatically reducing both training costs and storage overhead. The "fusion strength" (scale/weight) can also be adjusted at inference time to flexibly control how much the LoRA influences the output. In video generation, LoRA is commonly used to inject specific motion styles, camera behaviors, or visual characteristics into a base model without retraining the entire large model from scratch. CrossView-Warp LoRA uses exactly this mechanism to graft "cross-view transformation" capability onto the MiniMax H3 base model in a lightweight way.

ComfyUI Workflow

The author provides a complete example workflow (JSON file) that can be imported directly into ComfyUI. The workflow is located in the minimax-h3/crossview-warp directory of their ComfyUI-Workflows dataset repository. For users already familiar with the ComfyUI ecosystem, this means an essentially zero-barrier path to replication.

Custom Node

The key to making this work is a custom node called CrossViewWarp, open-sourced in the GitHub repository cseti007/ComfyUI-CrossViewWarp. Notably, this node is identical to the one used in the previous LTX version — meaning the node logic remains stable, with only the underlying model switching from LTX to MiniMax H3. The author has also included an explanatory video (recorded for the LTX version, but applicable here since the node is unchanged).

Why This Is Worth Paying Attention To

From a technical standpoint, CrossView-Warp addresses the problem of "camera freedom" in video post-production. Traditional video generation models fix camera behavior at generation time, whereas using a V2V approach to remap the viewpoint of existing clips — essentially warping to a new perspective — is a form of generative-model-based novel view synthesis. This has real practical value for short-form video creation, storyboard adjustments, and any scenario requiring multi-angle footage.

For the open-source community, the project's significance also lies in its portability. Node reuse across models, ready-to-use workflows, and publicly available training details — this "swap the model, keep the pipeline" design philosophy enables the community to rapidly iterate on similar capabilities using new base models. As next-generation video models like MiniMax H3 continue to emerge, lightweight LoRA-based adaptations like this are poised to become an important part of the video editing toolchain.

Novel View Synthesis (NVS) is a classic computer vision problem. Traditional methods rely on geometric information captured by multi-camera setups (e.g., depth maps, point clouds, or NeRF implicit fields) to reconstruct 3D structure and render from new viewpoints. Generative-model-based view remapping takes a different approach: rather than explicitly reconstructing 3D geometry, the model implicitly learns the rules of viewpoint transformation from large amounts of video data, directly "imagining" the pixels of a new viewpoint at inference time. This approach has low input requirements (no multi-camera rigs or depth sensors needed), but at the cost of geometric consistency — the model may hallucinate content in occluded regions. The practical performance ceiling of CrossView-Warp depends heavily on how deeply the MiniMax H3 base model understands 3D spatial relationships.

Getting Started

For creators who want to try this out, the recommended path is: first clone the CrossViewWarp custom node from GitHub and install it in ComfyUI, then download the LoRA weights from Hugging Face, and finally import the official example workflow to get the basic pipeline running. Since the tutorial video was recorded for the LTX version, the main difference when using the H3 version lies in the model loading step — node parameters and operational logic can be followed from the original video.

One caveat: the information in this article is based on the developer's release announcement. Real-world performance on actual footage, and robustness across different video types (such as complex motion or multi-subject scenes), still needs to be verified through hands-on testing. As an actively evolving community project, it's worth keeping an eye on the repository for future updates.