H3 HyperFlow P7 · LOW 0:4 Setup Only (T8 EXP)

The first pass of a P7 segment: four steps from the absolute start of the sampling grid, at low resolution, producing the coarse content that the HIGH pass will finish. H3 HyperFlow P7 · LOW 0:4 Setup Only builds that stage - binds the model, the conditions and the absolute interval - and hands you a sampler and a sigma table without running anything.

Inputs

prepared_phase - the LOW instance from One Phase Native Conditions. The phase carries the geometry, the encoded prompt and the original P7 motion/audio conditions, so the setup doesn't need to guess any of it.

model - an independently selected MODEL. In a full P7 run you'd typically have two branches off the same H3 base, each with its own content LoRA: one for LOW, one for HIGH. The pack is explicit that LOW and HIGH patches are separate and that the same goes for prompts, noise and effects. You can also run them with different HyperFlow adapters loaded per branch, since each branch has its own dedicated loader.

positive / negative - the paired conditions. If you're using Prompt Relay, wire these from Pair Relay to ONE Phase, which binds the projected events to both the MODEL and the CONDITIONING; the setup passes them through untouched, and its protected motion guides are not something you get to substitute.

Outputs

model (patched, ready for a guider - or for an EAV apply first), sampler and sigmas (the absolute 0:4 table: this stage is not starting from a fresh full schedule, it stops at step four of eight), source_av (the latent the sampler starts from), stage_context (typed, for the optional Stage EAV node), positive / negative, and report_json.

Why 0:4 rather than "four steps"

Because H3's HyperFlow route is trained on an eight-point time grid, and the two-pass recipe uses the first half and then the second half. Each pass takes the absolute interval it's responsible for: LOW runs 0:4, HIGH runs 4:8 from a re-noised, enlarged latent. Neither phase pretends it's the whole schedule, which is why the sigma output is described as absolute and why you shouldn't splice a core sigma node in here thinking you're being clever.

That design also predicts the tradeoff you're paying: step-reduced recipes always give something back, and running 4+4 rather than a full 8 at target size is a resolution-versus-fidelity bet, not a free lunch. The pack's own notes are careful not to claim the two-pass route is sharper or better-seamed than a single 8-step pass at final size.

Wiring

Prepare LOW phase → [Pair Relay to ONE Phase, optional] → LOW 0:4 Setup
  ├→ model → [External Stage EAV for This Phase, optional] → BasicGuider
  └→ source_av + sigmas + sampler → ONE Stage Sampler → Bind Completed LOW Result

Install

Manager → search 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

Restart ComfyUI fully, refresh the browser. requirements.txt is empty by design - the base nodes use only what ComfyUI already ships, so install can't touch your Torch/CUDA stack. Before a run you need H3 base weights in models/diffusion_models, the Qwen text encoder in models/text_encoders, video and audio VAEs in models/vae, the original non-pruned HyperFlow adapter in models/hyperflow/loras/, and the learned 3D upscaler in models/latent_upscale_models. Loading that HyperFlow file through a generic LoRA loader is not a shortcut the pack accepts - there's a dedicated loader input for it and the graph will reject a stray HyperFlow artifact in a content slot.

Notes and gotchas

Active stage effects in this pack assume CFG1. The example graphs get CFG1 by feeding the phase's positive conditioning into the negative socket, so if you want a real CFG number you also need real negative conditions - and the negative socket on the conditioning node mirrors positive rather than accepting anything else.

If a low-VRAM run dies here rather than in HIGH, don't be surprised: LOW is cheap in pixels but it's still a 33B joint audio-video model with a two-time HyperFlow patch applied, plus the text encoder. Reserve headroom, and if you hit the reproducible comfy-aimdo double-owner crash that shows up on this pack's long-video routes, try launching ComfyUI with --disable-comfy-compiler on an isolated instance - that's what the author's own probe runs do.