MiniMax H3 VDN Two-Pass Refine / VDN二采 (Advanced EXP/T8)

Every distilled model trades detail for speed, and MiniMax H3's 8-step VDN route is no exception - you get a clip in a fraction of the full-model time, but you can see where the shortcut cost you. The T8 pack's answer is a two-pass trick: run VDN's 8 steps at low resolution, upscale the latent with a learned video upscaler, then spend a few more steps refining instead of re-rolling the whole clip. MiniMaxH3VDNRefinePlanT8Advanced (display name: "MiniMax H3 VDN Two-Pass Refine / VDN二采") is the planner for that second pass. It does the fiddly bookkeeping - rebuilding high-resolution conditioning, reconciling audio, slicing the right sigma schedule - so you can wire up the refine sampler without doing the math yourself.

This is squarely an "Advanced EXP/T8" node. It arrived in pack v1.75.0 alongside four two-pass workflows, and the author's own docs are blunt that it's experimental: human review found two-pass and single-pass "about the same" on a 0.52MP test pair, and it's not guaranteed every input comes out sharper. If you want raw quality per generation, the single-pass route is still the safe default. You reach for this when VDN's speed is the point and you want the detail back.

How it works

The mechanism is a handoff between three states. It expects a model that came out of the pack's OpenVDN Composer (the MODEL carries an attachment receipt proving a VDN stage is actually configured - feed it a plain model and it refuses), plus two latents: the av_latent from the latent-reconcile node (the upscaled, high-res one) and the first_pass_latent - the actual output of your first sampler, used as a validation reference. It checks that batch, channels, frame count and audio length all survived the handoff, that the refine canvas didn't shrink, and that there's no NaN/Inf anywhere.

Then it replans execution and slices the tail of VDN's own trained stage-sigma grid - the last refine_steps + 1 sigmas, strictly descending to zero - so the refine restarts where the distilled model actually knows how to denoise, not on an arbitrary schedule. First-pass audio is retained by default; the report_json output tells you whether the audio matches the first pass and what your total NFE (base steps + refine steps) will be.

The inputs and outputs

• model (MODEL) - from MiniMaxH3VDNModelComposerT8Advanced. Non-negotiable; the node errors out if the composer receipt isn't there.
• av_latent (LATENT) - the reconciled, learned-upscaled latent for the second pass.
• first_pass_latent (LATENT) - the first sampler's output, for shape and audio validation.
• refine_steps (INT, default 4, min 1) - how many steps the refine pass gets. Must stay below the stage's base steps (so under 8 for the DMD8 route).

Outputs wire up exactly where you'd expect: model → BasicGuider, sampler and sigmas → your second SamplerCustomAdvanced, plus report_json for inspection. Add a fresh RandomNoise to that sampler - and don't bolt another Turbo/EMA LoRA onto the VDN branch. This node is built on the assumption that you won't; stacking a second speed LoRA usually just breaks the handoff or quietly degrades output.

Installing and running it

Same pack as everything else: ComfyUI 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 fully. This node adds no new pip dependencies - but you do need the OpenVDN model package from HuggingFace (t8star/Vdn-Minimax-H3-Comfy), the learned latent upscaler (minimax_h3_latent_upscaler_3d_fp16.safetensors), and FFmpeg on PATH for the save step. The cleanest path is opening one of the four prebuilt 2026-09-08_H3_OpenVDN_*_TwoPass_EXP.json workflows from examples/workflows/10-speed - they have all the handoffs pre-wired, and you only edit the shared prompt and duration.

Common issues

• "VDN refine requires the OpenVDN Composer MODEL" - you bypassed the composer node. Don't.
• refine_steps out of range - it must be 1 to (base steps − 1), so ≤ 7 on the 8-step route.
• "canvas must not shrink" / handoff shape errors - the upscale step between passes isn't set up right; re-check the latent-reconcile wiring.
• Red nodes on load - update ComfyUI, the frontend, and Manager, then restart; updating only this plugin is usually not enough.

If your audio dies or turns to noise on the second pass, check the workflow's sampler, scheduler, and shift settings before suspecting the node - and on a 16GB card, run one task at a time. That's the VRAM reality of H3 regardless of which pass you're on.