MiniMax H3 OpenVDN Runtime Audit (Advanced/T8)

OpenVDN is the pack's fast path: a distilled, 8-step (or 50-step Stage B) way of sampling MiniMax H3, built by grafting a large hybrid-attention branch onto the base model. Before you let a ComfyUI pack start patching a 33B model's internals, you might want someone to check that everything is where it claims to be. That someone is this node. And here's the neat bit - it audits a 4.28GB branch file and hundreds of adapter tensors without loading any of them. The whole check runs on file hashes, sizes, tensor counts, and metadata, so a mis-placed or corrupted model gets caught in seconds instead of after a 4GB load and a silent garbage render.

If that level of paranoia feels excessive, remember what's at stake: you're about to hand this branch to the Comfy ModelPatcher and have it own the attention layers of an H3 model. A wrong-version branch doesn't fail obviously - it just makes subtly wrong video. This node is the cheap failure point that prevents the expensive one.

What it verifies

The audit is read-only and never mutates anything. In order of importance:

• Exact asset identity - the linear_branch/model.safetensors (4.28GB, 800 tensors), the default adapter (416 tensors), and the turbo adapter (726 tensors) must match the pinned OpenVDN release's byte sizes and SHA-256s, or the pack won't touch them. verify_hashes defaults to true; turn it off only if you know exactly why.
• Native H3 structure - the base model must actually be the right kind of H3: the full base wants a 2688-wide AdaLN input, and if you're on the curve-pruned INT8 base the audit checks for the 8-wide AdaLN signature and matches the adapter to your base's curve table by content hash, not by filename.
• Attention ownership - OpenVDN takes over the model's attention, so this node rejects any competing owner (SLA, VSA, Sol-Attn, BlockCache, other model-patch nodes) that would fight it. "Just stack another accelerator" is how you get broken results; the audit calls it out before sampling.
• The base provenance boundary - if your base isn't the exact pinned upstream revision, allow_structural_base exists as an explicit, recorded exception. Leave it off unless the tooltip situation genuinely applies (like the local INT8/ConvRot base), and understand that enabling it is you accepting an "unproven, structurally compatible" verdict rather than relabeling it exact.
• License documents - the report lists the local NOTICE, source LICENSE, and MiniMax H3 Community License file, and reminds you the weight license is territory-restricted.

The inputs

model (the native H3 MODEL you loaded), vdn_root (default OpenVDN/vdn-minimax-h3 - where the branch lives under your diffusion_models folder), and stage (stage_dmd_8nfe default, or stage_b_50nfe). Outputs are the untouched model, a ready boolean, and a report_json that doubles as your audit trail. Connect ready into the downstream logic and check it's true before the Model Composer runs.

Installing and running it

Install the pack (ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes/) and restart fully. Then fetch the model package - the install-ready mirror is on HuggingFace:

hf auth login
hf download t8star/Vdn-Minimax-H3-Comfy --local-dir ComfyUI/models

That lands the branch and adapters under models/diffusion_models/OpenVDN/vdn-minimax-h3/ alongside your H3 base in diffusion_models. Nothing downloads at runtime.

Gotchas

• US/EU/UK/Korea, take note. The weights you're auditing sit under the MiniMax H3 Community License, which excludes those territories - review the terms before running anything, not after.
• ready = false with a filename complaint means the mirror got truncated or mixed. Re-download rather than fudging hashes - the pack's signature checks exist so you can't quietly run a mismatched adapter.
• Red workflow nodes after install almost always mean outdated ComfyUI core, not a bad pack - update ComfyUI + front-end + Manager and restart completely.