MiniMax H3 Chunked Two-Pass Global Noise Plan (v2 Advanced EXP/T8)

This node exists to fix a specific, real failure mode, so it helps to know what it's fixing. In the T8 pack's "learned latent two-pass" route you generate MiniMax H3 low-res and fast, push the clean latent through a learned 3D upscaler (models/latent_upscale_models), then re-sample the big canvas to add real detail. The problem: if the second pass runs in spatial tiles and temporal chunks that each start from their own noise, your "one" video comes out looking stitched - seams in space, and temporal trajectories that drift every time the model restarts from the previous chunk's output.

The v1 plan node handled that by chunking and carrying overlap. This v2 node is the opt-in correction, and the core idea is in its name: create one full target-resolution noise field for the whole video, once, and slice it by exact coordinates. Every piece of the second pass draws its initial noise from that single global field instead of rolling its own, which is what gives chunk boundaries something to agree on.

What the plan actually is

What you actually run this as: a plan node. It doesn't touch the sampler. It outputs a T8_H3_CHUNKED_TWO_PASS_PLAN (plan) plus a report_json audit string, and you feed the plan into the existing MiniMax H3 Chunked Two-Pass Upscale executor in place of its old v1 plan. The executor, the workflows, and v1 itself are all untouched - this is append-only, swap the plan wire and you're on v2.

The two switches that matter

Two combo inputs are the whole story, and both default to the safe route:

• spatial_strategy = full_frame_safe (default): one spatial canvas, no independent tiles. independent_tiles_exp exists but the author is blunt that separate spatial H3 Transformer tiles change the model's local coordinates and failed real quality review. Leave it alone.
• temporal_strategy = full_clip_safe (default): one complete H3 temporal trajectory, no stitching. guarded_overlap_exp inherits overlap from the previous chunk and still merges separate trajectories - it produced boundary spikes up to ~2.8× the frame-difference P95 in testing. Diagnostic only.

In the default route, the tiling widgets are basically dormant: tile_width/tile_height, spatial_overlap, spatial_fade, temporal_chunk_frames, temporal_overlap_frames and anchor_strength only bite when you opt into those EXP routes. The inputs a beginner actually touches are model_name (pick the learned upscaler - the pack expects minimax_h3_latent_upscaler_3d_fp16.safetensors in models/latent_upscale_models), target_width/target_height (the high-res canvas, multiples of 32), and maybe precision (fp16 default) if VRAM complains. release_policy defaults to offload_after, which is the right call on a 16 GB card.

Audio is handled safely by design: the per-piece audio noise stays zero, and the executor returns the exact original audio tensor. So this node is purely about not wrecking the video side of the joint AV latent.

Where people trip up

One workflow gotcha the author calls out explicitly: native H3 resizes the first endpoint and center-crops the last one. Feed the same non-square source straight to both ends and you get ~9% expansion mismatch in the render. The bundled workflow (13-latent-upscale/2026-08-30_H3_Chunked_TwoPass_Global_Noise_v2_Advanced_EXP.json) center-crops both endpoints to one shared canvas first - keep that wiring when you reuse it.

Installing it

Install the pack via ComfyUI Manager (search MiniMax H3 Audio T8) or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git into custom_nodes, then restart. Update ComfyUI core, frontend and Manager together first - this pack needs comfy_api.latest, comfy.weight_adapter, comfy.patcher_extension and comfy.ldm.minimax, so an old core makes every T8 node go red at once.

Honest verdict: on a card that can fit the full-frame pass, this is the two-pass plan to reach for. The tiled/time-chunked EXP options are traps the author already ran so you don't have to.