Meridian 1 · 模型与资源（INT8/T8）

What it is, and why it comes first

The Meridian chain is this pack's camera-move pipeline, and it runs in four steps: resources → material window → camera/time editor → three-forward generation. This node is step one - the config node - and it's the one that fails first if anything is wrong, which is exactly why you want it at the front.

Nothing is generated here. It resolves and validates every heavy artifact the other three nodes will touch: the converted Meridian checkpoint (34 GB, DMD already merged, native ConvRot INT8), the H3 video VAE, the pinned upstream Meridian source, and Meta's VGGT-Omega geometry stack. Upstream Meridian is Viggle's re-camera model; the converted file keeps the H3 legs - the H3 video VAE, the 17n+5 frame grid, the joint AV sampler - which is why the weights fall under the MiniMax H3 Community License while Omega carries its own FAIR non-commercial research licence.

The "audit" part is literal. Each Meridian source file is SHA-checked against the pinned revision, and the Omega directory must actually contain vggt_omega/__init__.py. Get those wrong and you get an exception in under a second instead of a confusing failure forty minutes into a decode. If you're coming from the depth/geometry side of ComfyUI, the shape of the dependency will be familiar: this is VGGT-class multi-view geometry, i.e. a model that recovers camera pose from views, not a grayscale depth map for a ControlNet.

The inputs that matter

Everything is a path, and every path has a sensible default, so a standard install is a lot of empty boxes.

• Leave the path fields blank and the node auto-finds models/meridian/meridian_dmd_int8_convrot_comfy.safetensors, models/vae/minimax_h3_video_vae_fp16.safetensors, models/meridian/source, models/meridian/vggt-omega, and the Omega weight under vggt-omega/checkpoints/. Env overrides exist if your weights live on another drive: MERIDIAN_MODEL, MERIDIAN_VIDEO_VAE, MERIDIAN_SOURCE_DIR, VGGT_OMEGA_DIR, VGGT_OMEGA_CKPT.
• cache_gib is a disk cache budget, default 20, and it is not VRAM admission. 0 turns the disk cache off. There's a matching cache_directory if you want it somewhere other than the default.
• The optional model / video_vae sockets are how you feed a MODEL or VAE you already loaded. Pick one route or the other: if you connect video_vae, don't also fill vae_path - the node raises instead of guessing which one you meant.

Outputs are runtime (T8_MERIDIAN_RUNTIME), which is the only thing node 2 accepts, and report_json - the same receipt trail the rest of the chain writes, useful when you want to know what actually got verified.

One detail worth knowing: this node always re-runs rather than being served from ComfyUI's cache, so swapping a weight file on disk re-triggers validation instead of inheriting a stale result.

Installing the pack and the models

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

Then fully restart ComfyUI and refresh the browser. ComfyUI Manager also finds it under the pack title MiniMax H3 Audio T8; the registry and GitHub release progress independently, so if you don't see the newest version, install from GitHub. The pack's requirements.txt is intentionally empty - base nodes use only what ComfyUI already ships, and EXP features check their own optional deps when used.

Weights come from t8star/Meridian-Comfy. Merge its models/ folder into ComfyUI/models/ rather than dumping every file in one directory. Grab the Meridian source and assets before merging the weight directory, or your git clone will hit a non-empty folder. And do not stack a DMD LoRA on top of this checkpoint - DMD is already merged into it.

Where it goes wrong

• "Correct authorized Omega source required, not ordinary VGGT." You pointed at regular VGGT, or at the 256 variant. Omega 1B512 is a specific download, and it isn't INT8.
• A rewrite of the pinned source revision fails the SHA check. Upstream moved; re-audit deliberately instead of mixing revisions.
• 34 GB is disk, not VRAM. This node doesn't move the main model onto the GPU - only the generation node does. File size tells you nothing about whether your card will finish the job.
• Licence boundaries are separate. Node source is GPL-3.0-or-later; Meridian's derivative weights follow the H3 community licence, whose applicable territory excludes the EU, UK, Republic of Korea and the US; Omega is non-commercial research use only. That's a licensing fact, not a bug you can troubleshoot.