VELVET VICE MiniMax H3 — Ghost Analyzer

If you've ever pushed a video to 48 or 60 fps with frame interpolation, you know the artifact: on fast motion, moving objects leave translucent after-images - "ghosts" - as the interpolator guesses frames between the real ones. It's the classic finishing-stage trade-off: interpolation makes motion smooth, and smoothness buys you ghosting. VelvetViceMiniMaxH3GhostAnalyzer is the node that finds those ghost frames in MiniMax H3 output so a repair stage can fix them.

This is a finishing-path node, and understanding that context is half the story. The typical open-weight video production chain - the KB's own standard pipeline for Wan-style work, and H3 follows the same shape - is: generate, optionally interpolate frames for smooth motion, upscale, add audio. Once you add that interpolation step, you've signed up for artifact inspection. The H3 Ghost Analyzer is the inspection half; its sibling TemporalAntiGhost node is the repair half.

How it works

The node takes the images (IMAGE) batch you've finished rendering and runs a temporal analysis over it. Its mode input tells it what it's looking at: SOURCE 24 FPS (the raw H3 output, which shouldn't have much ghosting) or RIFE INSERTED FRAMES (frames your RIFE-style interpolation inserted, which is where ghosts live). Then a stack of detection controls:

• profile - SAFE / BALANCED / STRONG, an aggression ladder for how eagerly it flags ghost regions.
• sensitivity (default 1.0), motion_search_radius (5), minimum_ghost_area (24) - how much motion it searches, and how big a ghost has to be before it counts.
• scene_cut_protection (default true) with a scene_cut_threshold (0.24) - so a genuine cut between shots doesn't get misread as a giant ghost region.
• analysis_long_edge (256) - it analyzes at reduced resolution for speed; ghost detection doesn't need full res.
• analysis_chunk_size (4) and memory_mode (AUTO/GPU/CPU) - chunking and where the analysis runs, because full-res frame batches are memory hogs.

Outputs are the tool's real payload: a ghost_analysis bundle that feeds straight into the TemporalAntiGhost repair node, plus ghost_mask (MASK) and mask_preview so you can actually see what it flagged, average_score and maximum_score floats for quantifying the problem, and an analysis_report (STRING) summarizing it.

Inputs and outputs that matter

• images (IMAGE) - the rendered frame batch.
• mode - SOURCE 24 FPS or RIFE INSERTED FRAMES. Getting this wrong mis-calibrates everything.
• profile - how aggressively to detect.

Outputs: ghost_analysis (→ TemporalAntiGhost), ghost_mask, mask_preview, average_score, maximum_score, analysis_report.

Install and how to use it

Same pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

then restart. It's a pure CV/tensor analysis node - no models to download, no heavy dependencies.

Where people get burned: they run it on RIFE-inserted frames but leave the mode at SOURCE, or they crank STRONG on a clip with real scene cuts and watch every cut get flagged as ghosting. Check the mask_preview before trusting the scores - if the mask is lighting up on intentional cuts rather than motion trails, adjust the scene-cut protection, not the sensitivity. And be realistic about scope: this analyzes and flags so the repair stage knows where to act. Analyzer alone doesn't fix anything; it's the first half of a two-node finishing pair.