Meridian 3 · 运镜／时间编辑器（T8）

Camera moves on generated video are usually guesswork: type a word like "orbit" into a prompt and hope. The Meridian route in this pack does something different - it reconstructs actual scene geometry from your material and lets you author the camera path in a spatial track and a source-time track, then previews the warp before any diffusion happens.

MiniMaxH3MeridianCameraEXPT8 is the third of four Meridian nodes (Config → Material → Camera → Generate). It's the editor, and it's the one that turns a plan into a prepared payload.

What it takes and gives back

Inputs: material (from the Meridian material node), frames (default 73, minimum 22), preset (default slide), strength (default 0.08), and camera_plan - a JSON field the big editor writes for you. Leave camera_plan empty and the preset is used instead.

Presets available: slide, push_in, pull_out, crane, orbit, freeze_orbit, source_camera. strength is displacement relative to a fixed pivot depth; for orbit it's multiplied by 100 degrees. The tooltip - worth internalising - says larger displacement means more hole-filling, and explicitly refuses to promise performance or quality.

Outputs: prepared (the payload the generate node consumes), canonical_plan (the exact JSON contract, so a saved plan is reloadable and API-readable), warp_preview (an IMAGE), and report_json. The report carries coverage numbers - min and mean - which is a genuinely useful signal, because a warp that has to invent large areas tells you so numerically.

The preview is the point, and it isn't a render

Run a preset first and look at the geometry preview. The author's own description is careful: it's a grey-hole point-cloud warp, not generated video. Those grey holes are the regions where the camera move exposed nothing in the source, and no diffusion pass has filled them yet. Reading the preview as "this is what the shot will look like" is the classic misunderstanding of this whole node.

The rest of the contract is fixed and stated plainly: first-source camera coordinates, fixed depth units, roll = 0, no reverse source time, and no automatic character tracking. The spatial and source-time tracks are independent. If you like orbiting a subject while keeping it centred, understand that slide explicitly does not lock the subject - the camera and look point translate together with constant orientation, as a parallel dolly. That was a deliberate fix: an earlier, softer version of the preset kept the subject centred and got feedback that the motion was invisible.

The one rule you cannot skip

Change the material or the frame count and you must explicitly reset and re-edit the path. The node refuses to run a saved plan whose frame count no longer matches the output, with an error telling you to reset or update both tracks in the editor. There's no silent migration of old paths, on purpose - and hand-written or saved canonical plans are never rewritten by preset changes.

Install

ComfyUI Manager → MiniMax H3 Audio T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Full restart of ComfyUI, then refresh the browser - this one installs a frontend editor, so a page refresh alone won't surface it. The pack's requirements.txt adds no packages; the Meridian route's own dependencies are model files, and they are not small. From t8star/Meridian-Comfy: the converted merged-DMD native ConvRot INT8 model (about 34 GB) to models/meridian/, the pinned Meridian source and paired-length assets/ to models/meridian/source/, the VGGT-Omega source to models/meridian/vggt-omega/, and the correct Omega 1B512 checkpoint to models/meridian/vggt-omega/checkpoints/ - plus the H3 video VAE in models/vae.

Two licensing notes while you're downloading: the Meridian derived weights follow the MiniMax H3 Community License (the grant is void in the US, EU, UK and South Korea, outputs included), and Omega carries its own FAIR non-commercial research licence. They don't merge.

Where it bites

DMD is already merged. Don't stack a DMD LoRA on top of the converted model - the pack says this in bold in the README, and it's the kind of thing that quietly wrecks your output rather than erroring.

roll is not editable. It's fixed at zero. If your shot needs Dutch angle, this isn't the tool.

Bigger windows can OOM. The material node's geometry work happens once over the selected window, and large windows can exhaust memory. It won't auto-shrink or silently retry.

Geometry truth isn't output quality. A measured, frame-stable camera orientation proves the warp does what the plan says - not that the filled regions hold up.

The four-node chain is the unit. This node generates nothing by itself; a preview and a prepared payload on their own are not a clip.