MiniMax H3 Tiled VAE Coordinate Audit (T8 Advanced EXP)

Big H3 canvases eventually hit a wall: the VAE decode won't fit in VRAM, so people reach for tiled VAE - decode the frame in chunks and stitch. Tiled decoding works, but it has a telltale artifact: each tile carries its own coordinate context, and the seams between tiles can show up as faint grid lines or brightness discontinuities. The community's proposed fix was "global coordinates" - tell every tile where it sits in the whole image so the model's positional encoding is consistent across the grid. Sounds right, and this node is the record of what happened when T8 actually tested it on the real fp16 H3 VAE.

The result was not what the proposal hoped. On H3's actual fp16 VAE, global coordinates regressed - they made the regular grid artifacts worse instead of better. So the pack's verdict, encoded in the node's default behavior: report only. It audits the upstream proposal, reports what it finds, and returns your source VAE completely unchanged. There's an experimental apply mode in the schema, but the README is explicit that the global-coordinate candidate is kept as a bypassed experiment and is not recommended as a fix. This is a node that exists partly to document a negative result.

How to use it

• video_vae in → video_vae out (unchanged in the safe default)
• mode - report_only (default) or apply_global_coordinates_exp
• report_json - what the audit found

Wire the VAE through it, read the report, and move on. If you were already happy with plain tiled decode, nothing here changes your pipeline - that's the design.

Installing it

Part of the T8mars/comfyui-minimax-h3-audio-T8 pack - search "MiniMax H3 Audio T8" in ComfyUI Manager, install, restart, or:

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

Update ComfyUI itself first (recent core with comfy_api.latest, comfy.weight_adapter, comfy.ldm.minimax), and don't pip-install extras - requirements.txt is intentionally empty.

The honest take

This is a "we checked, it doesn't work, here's the audit" node, and that's genuinely useful as documentation - it saves you from burning an afternoon re-testing a plausible-sounding fix. But don't go hunting for apply_global_coordinates_exp expecting a hidden win; the pack's own measured result on the fp16 VAE says no. If you're seeing grid artifacts in H3 output, your first stops should be the boring ones: check your VAE isn't a mismatched variant (the README's "don't mix FL2VA/Ref2VA/pruned/full" rule), confirm width/height are multiples of 32, and drop resolution before assuming a VAE-coordinate fix is the answer. Tiled VAE stays the escape hatch for frames that don't fit - just expect it to cost some seam quality at the edges, which is the trade you were already making.