- family
- 3d
- modality
- video -> per-frame geometry -> depth-warped guide video -> re-angled video
- whyMultiModel
- Meridian ships as LoRA adapters on the unmodified MiniMax-H3 transformer, not as its own checkpoint, so on its own this would be a single generative backbone with a LoRA, the same trap that sank the quarantined FLUX.2 Klein 9B consistency-LoRA candidate. The multi-model claim instead rests on MoGe-2, a separately trained monocular geometry model the video transformer cannot substitute for: it produces the per-frame point-cloud geometry that Camera H3 reprojects into the depth-warp guide Meridian conditions on, and Meridian cannot derive that geometry itself. This is the same accepted shape as the tracked depth-pose-guided-v2v-restyle and controlnet-structure-lock patterns (an annotator/geometry model feeds a conditioned video diffusion model), not a from-scratch second checkpoint, so confidence here is set accordingly. It is distinct in purpose from those tracked patterns and from depth-video-character-reskin-seedance and worldmirror-splatguide-pose-free-novel-view: depth-pose-guided-v2v-restyle changes appearance/style while holding the camera fixed; depth-video-character-reskin-seedance swaps the on-screen performer while holding the camera fixed; worldmirror-splatguide needs multiple unposed input views and outputs still frames. This chain instead moves the camera itself (orbit, pan, zoom, dolly) and reframes time from a single existing video, with the warp guide, not the text prompt, driving the new camera motion.
- steps
- [object Object], [object Object], [object Object]
- controls
- Virtual camera path preset (orbit/pan/zoom/dolly) authored in Camera H3; temporal controls (slow down, pause, resume) independent of the spatial camera path; Quality INT8 versus LoRA H3 Base workflow variant; DMD/3-step LoRA acceleration; MoGe geometry checkpoint choice.
- exampleStack
- MoGe-2 (geometry) -> Camera H3 (depth-warp reprojection) -> Viggle Meridian LoRA on MiniMax H3 (re-angled video generation).
- useCases
[Re-angling single-take footage into new cinematic shots (orbit, dolly, push-in) without reshooting][Virtual re-cinematography of archival or UGC video that was only ever captured from one camera][Combined spatial and temporal reframing (new angle plus slow-motion or freeze) of an existing clip]
- pitfalls
- Regions disoccluded by the new virtual camera angle have to be filled in by Meridian rather than truly reconstructed, so extreme viewpoint changes or thin/occluded geometry in the source video degrade the fill; because the camera path is derived entirely from the MoGe depth warp rather than the text prompt, a noisy or low-confidence depth estimate propagates directly into warped, unstable geometry in the final render.
