<h2 id="stubelius-director-minimax-h3-seed-hunt-+-learned-2-refine">Stubelius Director — MiniMax H3 Seed Hunt + Learned 2× Refine</h2>A complete hunt → pick → refine pipeline for MiniMax H3, built on the Stubelius Director node pack — my fork of Muse Collective's excellent Muse Minimax Director V1.2, rebuilt around a trained 2× latent upscaler, a take-preserving audio lock, and a redesigned two-stage architecture that fixes the problems we measured in the original.Node pack (GitHub): <a target="_blank" rel="ugc" href="https://github.com/stuubszzz/Stubelius-Director">https://github.com/stuubszzz/Stubelius-Director</a> <div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/JAdoKVSE2LY" start="0"></iframe></div><h3 id="what-it-does">What it does</h3>Write your script as CUTs on a visual timeline, drop in character reference images / video / audio, and scout multiple seed candidates cheaply. Preview the candidates, press one button on the Refine dashboard, queue again — only your pick gets the expensive high-resolution finish. Seed, steps, split point, sampler, scheduler, and the compiled prompt all travel embedded inside each candidate latent, so the Refine auto-syncs to the exact candidate you picked. Nothing to match by hand.<h3 id="why-the-upscaler-was-replaced">Why the upscaler was replaced</h3>The original Refine upscales the mid-schedule latent with nearest-exact interpolation — its own docs cap the stable range at ~1.3–1.5× and warn of artifacts beyond. Interpolation invents no detail. Stubelius swaps that step for the trained H3 clean-latent 2× upscaler (Mamad8's model, via Tr1dae's integration): a neural network trained specifically to double H3 video latents, making a clean 2.0× the standard operating point instead of the danger zone.<h3 id="why-the-two-stage-pipeline-was-restructured">Why the two-stage pipeline was restructured</h3>We A/B-tested the original V1.2 two-stage against full single-pass hunts and isolated its core flaw: it upscales and continues from a half-baked mid-schedule latent. Consequences we measured: candidate previews that don't match outputs, motion timing drifting off the music, and audio being audibly regenerated — the take you auditioned never survived. Stubelius fixes the sequencing: complete the candidate's own trajectory first, then upscale the clean result, then polish — the upscaler receives what it was trained on, and audio finishes through the full schedule before being frozen.<h3 id="the-three-modes">The three modes</h3>⚡ Fast Hunt (two-stage, fixed) — Two-Stage ON + Latent-Only Scouting ON, first-pass 3–5 steps. Candidates cost only those steps. The Refine's <code>complete then polish (stubelius)</code> strategy then: (1) finishes your pick's remaining schedule at scout resolution — the exact trajectory the preview estimated, audio completing to full quality, (2) learned-2× upscale of the clean take, (3) tail polish with the audio lock. Roughly 3× cheaper per hunt cycle than full hunts. Trade-off: previews are rough estimates — audition composition and motion, not final detail or audio.🎯 Trustworthy Hunt (complete candidates) — Two-Stage OFF, full steps (e.g. 20). Every candidate is a complete generation: what you see and hear is real. The Refine auto-detects complete candidates and runs the polish path directly: learned 2× + a <code>refine_denoise</code> tail — with the exact audio you auditioned frozen into the final. Use this when the take (especially audio) must be auditioned before committing.🔧 Stock — the original V1.2 mid-schedule continuation, still selectable (<code>continue mid-schedule (stock)</code>), for comparison or multi-chunk bundles (which always use stock).<h3 id="the-audio-lock">The Audio Lock</h3>The original Refine regenerated audio every time — a side effect of continuing from the raw half-denoised audio intermediate. Stubelius' <code>audio_mode: keep candidate audio (locked)</code> freezes the finished audio with a zero noise-mask on the audio stream: pass 2 re-samples video only, and your final's soundtrack is exactly the take that came out of the hunt/completion. Switch to <code>continue schedule (stock)</code> when you want audio re-rendered against the upscaled video.<h3 id="refine-dashboard-controls">Refine dashboard controls</h3><ul><li>Candidate 1–4 — one-button pick from the hunt</li><li><code>sync_from_director</code> (on) — pulls seed/steps/split/sampler/scheduler/prompt from the candidate itself; the widgets below it become fallbacks</li><li><code>two_stage_strategy</code> — <code>complete then polish (stubelius)</code> vs stock continuation</li><li><code>refine_denoise</code> — how much the 2× polish may change the take: 0.3–0.35 maximally faithful, 0.4 default, 0.5+ cleaner but freer</li><li><code>polish_steps</code> — how carefully it polishes: 0 = auto-slice; 12–20 = a dedicated schedule over the same noise range (more convergence, identical faithfulness — recommended for fast/complex motion)</li><li><code>audio_mode</code> — the lock, see above</li><li>Upscale method — <code>learned model (gold, 2x)</code> (factor slider ignored; always exactly 2×); interpolation methods remain as dependency-free fallbacks</li></ul><h3 id="timing-tips-(avoid-off-beat-video)">Timing tips (avoid off-beat video)</h3>H3 generates frame counts only on a 17k+5 grid. A 13s request silently becomes 13.67s (328 frames) while a 13s reference clip trims down to 12.96s (311) — that mismatch alone drifts your video off the music, in the first pass. Pick durations on the grid (12.9s → exactly 311 frames) and trim reference audio to match. And remember: H3 generates audio guided by your reference — a stylistic reproduction, never a sample-accurate copy. For beat-perfect music videos, mux your original track over the final.<h3 id="required-custom-nodes">Required custom nodes</h3>PackWhyInstall<a target="_blank" rel="ugc" href="https://github.com/stuubszzz/Stubelius-Director">Stubelius-Director</a>Director + Refine<code>git clone</code> into <code>custom_nodes</code><a target="_blank" rel="ugc" href="https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler">ComfyUI-MiniMaxH3_LatentUpscaler</a> (Tr1dae)Learned 2× integrationManager<a target="_blank" rel="ugc" href="https://github.com/mamad8c/ComfyUI-H3-Latent-Upscaler-Mamad8">ComfyUI-H3-Latent-Upscaler-Mamad8</a>Upscaler architecture + checkpoint (auto-downloads)ManagerKJNodesGet/Set wiring + previewsManagerVideoHelperSuiteVideo outputManagerComfyUI-LayerStylePurgeVRAM (optional)ManagerBoth upscaler packs are required (Tr1dae depends on Mamad8 as a sibling). Needs recent ComfyUI core with native H3 support and <code>pip install av</code>. Models: H3 ref2va checkpoint (+ fl2va for First/Last Frame mode), matching CLIP, video VAE, audio VAE — from the official <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3">MiniMax H3 repo</a> under its license. Update: 24.8.2026: Major fix: if the Refine silently did nothing (queue finishes in 0.01s, no output) — that was un-run candidates blocking the node when fewer than 4 candidates were hunted. Candidate inputs are now lazy: only your picked candidate is evaluated, so all 4 wires can stay connected with any number of candidates ticked. Also fixed a crash in complete-then-polish, <code>polish_steps</code> validation errors on older saves (now self-healing, default 16), and speaker-chip / drag-highlight feedback.<h3 id="quick-start">Quick start</h3><ol><li>Load the included workflow. Drop character refs, write your CUTs, set an on-grid duration (12.9s).</li><li>Pick your mode: fast hunt (Two-Stage ✓ + Latent-Only ✓, first-pass 4, steps 20) or trustworthy hunt (Two-Stage ✗, steps 20).</li><li>Queue → review candidates → press <code>1–4</code> on the ⭐ Stubelius Refine → queue again.</li><li>Console confirms the pipeline: <code>complete-then-polish: finished remaining N steps at 1x</code> → <code>[polish] dedicated schedule</code> → <code>effective 2.000x/2.000x</code> → <code>audio lock: candidate audio frozen</code>.</li></ol>Ships with a full black &amp; neon-gold graph theme — every node, wire, and dashboard.<h3 id="credits">Credits</h3><ul><li>Muse Collective — the original <a target="_blank" rel="ugc" href="https://github.com/muse-collective-26/MiniMaxH3-Director-V1.2">Muse Minimax Director V1.2</a> (MIT): the timeline UI, chunking, seed hunt, and continuity engineering are their work.</li><li>Mamad8 — trained the H3 clean-latent 2× upscaler · Tr1dae — the ComfyUI integration this fork calls.</li><li>Powered by MiniMax H3 — <a target="_blank" rel="ugc" href="https://huggingface.co/MiniMaxAI/MiniMax-H3">open-weights model</a>, used under the MiniMax Community License.</li></ul>

Stubelius Director — MiniMax H3 Seed Hunt + Learned 2x Refine/Upscale Workflow

Improved stability and bug fixes Purge cache nodes after every seed &amp; Model purge: False (before:True) NOTE: Using nodes 2.0 with Rgthree powerlora loader may break parts of the workflow.

Files with SHA256: 33358c25eb4051dc51ad19fd024ce3fbeba08ee7247bcddca57a4a37a5a2fce8