# MiniMax H3 Preview OverrideLive, kjnodes-style sampling previews for MiniMax H3 in ComfyUI — with true RGB colors.MiniMax H3 is an omni-modal model (video / audio / text / image). Like most videomodels it has no tiny autoencoder bundled, so ComfyUI's built-in sampler previewcan only show a rough Latent2RGB projection — which, for a 24-channel video latent,looks like colored noise.This single node attaches an <code>OUTER_SAMPLE</code> wrapper to the H3 model and, **on everysampling step**, decodes the current <code>x0</code> video latent into real RGB frames andstreams them into a DOM widget right on the node — with sigma / delta graphs,per-step timing and a click-to-scrub history.It ships with a trained Tiny AutoEncoder decoder <code>taeh3_decoder.safetensors</code>,24-channel, ~9.9M params) that distills the big H3 video VAE's decoder — the H3equivalent of <code>taeltx</code> for LTX or <code>taesd</code> for SD.---## Features- 🔴 True-RGB per-step previews — real decoded frames, not Latent2RGB noise- ▶️ Animated previews — sample multiple frames per step, play back asanimated WebP (or H.264 MP4 when NVENC is available)- 📊 σ / Δ graph — hover to scrub steps, click to lock and inspect any step- ⏱️ Step timing — per-step ms and rolling average + ETA in the widget header- 🪶 Cheap — the TAE is ~40 MB and decodes in milliseconds; negligible VRAM- 🛡️ VRAM-safe — only tiny autoencoders (TAEHV/TAESD) are ever used per step;a full video VAE is ignored (it would OOM during sampling)- 🖼️ Full resolution — previews stream at sampler resolution (capped by<code>max_resolution</code>), unlike the default 512 px first-frame preview- 🧹 Clean logs — the expected "missing encoder keys" warning for decoder-onlyTAE files is suppressed## Requirements- ComfyUI (current 2025+ version with the new frontend and <code>comfy_api.latest</code>)- The [MiniMax H3 model pack] (<a target="_blank" rel="ugc" href="https://www.minimax.io/blog/minimax-h3">https://www.minimax.io/blog/minimax-h3</a>) — the<code>MiniMaxH3ReferenceToVideo</code>, <code>MiniMaxH3SigmaShift</code>, sampler and VAE nodes- ~24 GB+ VRAM for H3 video generation (the model itself is ~20 GB)- Python <code>av</code>/PyAV is optional (only needed for NVENC MP4 previews; WebP worksout of the box)## Installation[code]cd ComfyUI/custom_nodesgit clone <a target="_blank" rel="ugc" href="https://github.com/simsim9-stack/ComfyUI-MiniMaxH3-PreviewOverride.git">https://github.com/simsim9-stack/ComfyUI-MiniMaxH3-PreviewOverride.git</a>[/code]Or download the ZIP from GitHub and extract it into <code>ComfyUI/custom_nodes/</code>.Then:1. Restart ComfyUI (the node registers on startup).2. Hard-refresh the browser tab (Ctrl+F5) so the widget JS is loaded.### Install the mini VAE (required for real previews)Copy the trained decoder into your <code>vae_approx</code> folder:```bash# from this repo:cp minivae/taeh3_decoder.safetensors ComfyUI/models/vae_approx/```Without this file the node falls back to Latent2RGB (the "colored noise" previews).## Usage1. Add *<code>MiniMax H3 Preview Override</code>** to your graph.2. Wire the H3 model into it (e.g. after <code>MiniMax H3 Sigma Shift</code>).3. Wire its MODEL output into your sampler's chain (guider + scheduler).4. In the <code>vae_name</code> combo select <code>taeh3_decoder.safetensors</code>.5. Queue. The widget on the node shows the video forming, step by step.### Node inputs| Input | Type | Default | Description ||---|---|---|---|| <code>model</code> | MODEL | — | MiniMax H3 model to attach the preview wrapper to || <code>tiny_vae</code> | VAE (optional) | — | TAE wired externally; leave empty to use <code>vae_name</code> || <code>max_resolution</code> | INT | 1024 | Max preview side in px; 0 = full sampler resolution || <code>jpeg_quality</code> | INT | 80 | JPEG/WebP quality of the streamed previews || <code>suppress_default_preview</code> | BOOL | true | Hide the default sampler preview overlay while sampling || <code>preview_frames</code> | INT | 8 | Frames sampled per step; 1 = single frame (fastest) || <code>preview_fps</code> | INT | 12 | Playback FPS for animated previews || <code>vae_name</code> | COMBO | taeh3_decoder.safetensors | TAE file from <code>models/vae_approx</code> |Output: MODEL — the model with the preview wrapper attached.### Widget- Header: resolution · step/total · avg step time · ETA- σ / Δ graph: sigma schedule with the per-step Δ (change in the latent) —hover to scrub through captured steps, click to lock a frame- step time graph: rolling per-step timing## Example workflowsThe <code>examples/</code> folder has the three base MiniMax H3 modes, each with thePreview Override node already wired between the model and the sampler:| File | Mode | What's different ||---|---|---|| <code>minimax_h3_t2v_preview_override.json</code> | T2V — text to video | no reference inputs || <code>minimax_h3_i2v_preview_override.json</code> | I2V — image to video | one <code>LoadImage</code> → first-frame reference || <code>minimax_h3_r2v_preview_override.json</code> | R2V — reference to video | two <code>LoadImage</code> → character/scene references |Common chain (all three):```UNETLoader → Sigma Shift → [MiniMax H3 Preview Override] → BasicGuider ─┐CLIPLoader ─┐ ├─→ SamplerCustomAdvanced → VAEDecode → CreateVideo → SaveVideoVAELoader ──┴→ MiniMaxH3ReferenceToVideo → conditioning + latent ───────┘```For R2V, reference videos and audio can be wired into the same node's<code>ref_videos</code> / <code>ref_video_audios</code> inputs — the prompt then refers to them as<code>&lt;Video k&gt;</code> / <code>&lt;Audio j&gt;</code>.Load a file via Workflow → Open, pick your model files in the combo boxes<code>unet_name</code>, <code>clip_name</code>, <code>vae_name</code>, audio vae), drop your own images into the<code>LoadImage</code> nodes, and queue.## How it works- An <code>OUTER_SAMPLE</code> wrapper replaces the sampler's step callback. On each step thepacked NestedTensor is unpacked to the video stream <code>[B, 24, T, H, W]</code> and decoded:1. Tiny VAE — <code>vae.decode()</code> of the TAE (true RGB, preferred)2. Animated Latent2RGB (multi-frame, when no TAE)3. Single-frame Latent2RGB (fallback)- The node validates the TAE's latent channels (H3 needs 24). A mismatched TAE(e.g. an LTX or Wan TAE) is disabled with one clear warning instead of erroringevery step.- Encoding (JPEG/WebP/MP4) happens on a background thread so the sampler neverblocks; full VRAM VAEs are never used for previews.## About the mini VAE<code>taeh3_decoder.safetensors</code> is a decoder-only Tiny AutoEncoder (TAEHVarchitecture, 24 latent channels, fp32) trained by the author on pairs of(raw H3 latents, decoded frames) generated with the official H3 video VAE —distilled from ~100+ H3 video generations. Holdout PSNR ≈ 25.5 dB.It loads through <code>comfy.sd.VAE</code> straight into the <code>TAEHV(latent_channels=24)</code>branch. Because it is decoder-only, ComfyUI logs a <code>Missing VAE keys [...]</code>warning on load — that is expected and harmless (the node suppresses it).## Troubleshooting| Symptom | Cause / fix ||---|---|| Preview shows colored noise | TAE not loaded — check <code>vae_name</code>, put <code>taeh3_decoder.safetensors</code> in <code>models/vae_approx/</code> || <code>Missing VAE keys</code> warning in the log | Expected for decoder-only TAE files; harmless || "Loaded TAE has N latent channels but the model needs 24" | You picked a TAE for another model (LTX/Wan/SD) || No widget / blank node | Restart ComfyUI and hard-refresh the browser (Ctrl+F5) || Widget shows only a single frame | <code>preview_frames = 1</code> — set it to 8+ for animated previews || OOM during sampling | Do not wire a full video VAE into <code>tiny_vae</code>; only tiny decoders are safe |## CreditsInspired by [KJNodes'](<a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">https://github.com/kijai/ComfyUI-KJNodes</a>)Model Preview Override, adapted for the MiniMax H3 video latent.MiniMax H3 by [MiniMax](<a target="_blank" rel="ugc" href="https://www.minimax.io/">https://www.minimax.io/</a>).

Files with SHA256: 200b17f16fbdf2afbd4f5c70b8390d57225bd2671ec17dfe162ad0e866dff66c