Inspect H3 RefMod
Read a mod's vitals before you blame the generation
- mods
- vae
- details
- image_preview
- audio_preview
"Throw the mod in and pray" works until it doesn't, and when a RefMod underperforms you need facts, not vibes. Inspect H3 RefMod is the pack's debug/inspection node: it reports exactly what a bundle contains - paths, latent shapes, token counts, concept type, description, and any config that Fix H3 RefMod Config baked in - and can show you a VAE-decoded preview of what's actually stored. Every other node tells you what it intends; this one tells you what's there.
How it works
Feed it a mods bundle and it returns a JSON details string (per-mod: name, path, kind, latent shape, token count, strength, description, concept_type, saved config) plus a total token figure for the whole bundle after copies. That total is the number to compare when you're asking "why is this run so slow" - token cost is the pack's main performance dial, and Inspect puts the real number in front of you.
The interesting part is the optional preview, which needs a matching VAE connected:
- preview =
off- details only (default). No VAE required. stored- VAE-decodes the mod as stored. Visual mods show the first latent frame; audio mods preview up to two seconds (requires the H3 audio VAE, and the node errors politely if you hand it the wrong one).compare_strength- decodes the stored mod and the same latent weakened bystrength, appended after it, so you can see how much a strength dial is actually blurring the reference.
index picks which mod in the bundle the preview uses. Image output feeds a normal preview node; audio output feeds an audio preview/speaker.
The README's honesty caveat belongs in your head here: this shows stored information, not a prediction of how well identity or style will transfer. A mod that previews beautifully can still fail at generation, and one that looks compressed can work fine in the right workflow. Previews are a reality check on the file, not a quality guarantee.
Install
ComfyUI Manager (search "ComfyUI-MiniMaxH3Mod") or:
cd ComfyUI/custom_nodes
git clone https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod
Restart; deps are light. You need H3 weights + its VAE(s) in a current ComfyUI for previews to decode anything, and remember H3's licence is territory-restricted (US, EU, UK and Korea excluded from local weights).
Common issues
The two failure modes are both VAE-related and both loud: connect no VAE with preview on and it raises "Connect the matching H3 VAE"; connect the video VAE to an audio mod's preview and it raises a type mismatch instead of silently decoding garbage. If details shows a config you don't recognize, that's a mod carrying baked settings from Fix H3 RefMod Config - check override on Apply before assuming the widgets are what ran. And keep expectations grounded: a preview that looks soft at a tiny grid isn't a bug, it's the documented pooling trade of training mode.
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| mods | H3_REF_MODS | — | |
| index | INT | 00–10000 | — |
| preview | COMBO | 3 options: off, stored, compare_strength | |
| strength | FLOAT | 0.500–1 | — |
| vaeopt | VAE | — | |
| visual_previewopt | COMBO | first_frame | Full video decodes every stored latent frame; increases memory use. Comparison appends the weakened sequence after the stored sequence. Audio remains limited to two seconds. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| details | STRING | — |
| image_preview | IMAGE | — |
| audio_preview | AUDIO | — |