johndpope/h3-neuron-explorer

H3 neuron explorer data

Data for the neuron explorer for MiniMax H3: where 1,333 characters and real people live in MiniMax H3, in its prompt encoder (Qwen3-VL-32B, the 50 layers H3's DiT reads; as a face and as a voice) and in its video DiT (read on H3's own clips of each character), plus each character distilled to one small file. With this download the explorer's characters tab works without any build step.

Credit

The character table and the test clips are from malcolmrey's malcolmrey/various dataset (h3-center/), released under the WTFPL and copied here with thanks. Find malcolmrey at https://huggingface.co/malcolmrey, https://reddit.com/r/malcolmrey and https://buymeacoffee.com/malcolmrey.

Contents

path	what	size
h3-center/crossovers/docs/CHARACTER_USABILITY_INDEX.md	738 characters: show, character, actor, usability with the base model (malcolmrey)	90 KB
h3-center/known-characters/	H3's own test clip(s) per character, 1,358 clips of 1,288 characters, INDEX.md maps names to clips (malcolmrey)	0.9 GB
h3-center/known-characters/prompts.jsonl	per clip: character, actor, show, status, the full H3 prompt, seed, sampler, steps, size and length, read from the ComfyUI workflow embedded in each clip	2 MB
thumbs/<name>.jpg	a 160 px frame (1 s in) of each character's first test clip; the explorer's picker thumbnails	6 MB
maps/sweep.pt	face map: per character, layer and unit, the mean MLP input (down_proj) at the character's name tokens; A [1333, 50, 25600] fp16, mu, sd, rows	3.4 GB
maps/voice.pt	voice map: a generic person says a line [English in X's voice …]; the same, measured on the spoken line's tokens after the voice tag	3.4 GB
maps/emb.pt	per character, the mean encoded prompt minus a neutral prompt (d), used for persona neighbours	27 MB
dit/clips/<clip>.safetensors	the DiT read on each test clip (below): delta_video, delta_face (named − name-removed), std_video and mean_video (the named pass), each [5 step windows, 50 blocks, 14336] fp16; the clip, both prompts, seed, sigmas, face box and audio in the metadata	~39 GB
dit/contrast/<character>_proxy_s4401/	for the 45 characters without a test clip (and a few others): a generated named-vs-neutral DiT contrast, z.pt [4 steps, 50, 14336] and meta.json with the top units and the patch / ablate test against random units	6 MB each
dit/population.safetensors	per window, block and unit: the mean and std over characters of the named − neutral shift (what a character's z is measured against)	29 MB
characters/<name>.safetensors	each character distilled (below)	~50 KB each

The character table is the 738-row usability index plus every character that has a test clip but no index row (595, usability from the clip's folder: good / onthefence / bad). The explorer builds it from the two files above.

The DiT, read on H3's own clips

H3 is a flow model, so its DiT can be read on a finished clip without sampling: 73 frames of the clip (from 1 s in, at its native 960 × 544) and its sound are encoded with H3's video and audio VAEs, noised with the clip's seed to the sigmas of steps 2, 6, 10, 14 and 18 of the 20-step schedule (one per step window 1–4 … 17–20), and the DiT (minimax_h3_fl2va_pruned_int8_convrot, the model that made the clips) runs one forward per noise level, twice: under the clip's prompt and under the same prompt with the character taken out (names, actor, show and voice tag replaced by "the person"). Same latent, same noise: the only difference between the two passes is the name. A DiT unit is one of the 14,336 values entering a block's mlp.fc2; per unit we keep the mean over the video tokens and over the face box (insightface on the clip).

The 45 characters with no test clip get a generated contrast instead (dit/contrast/): H3 renders the character's prompt and the same prompt with "a person" (proxy: 4 steps, 256 × 384, 22 frames, seed 4401), z per DiT unit between the two, then the top 8 / 32 / 128 units patched into the neutral run and ablated in the named one, against random units.

Check (Ace Ventura): the VAE round trip of the clip is 0.98 ArcFace to the real frames; the DiT's denoised prediction climbs from 0.03 (step 2) to 0.83 (step 18), so the latent and the noise are what the model expects.

A character in one file (characters/*.safetensors, format h3-character/1)

tensor	shape	what
dit_video_idx, dit_face_idx	[5, k] int32	per step window, the top-k DiT units by |z| (block × 14336 + unit)
dit_*_shift	[5, k] fp16	what naming the character shifts each unit by (activation units, averaged over its clips)
dit_*_z	[5, k] fp16	that shift against all characters (population.safetensors)
enc_face_*, enc_voice_*	[k_enc]	its top encoder units (layer × 25600 + unit), z and activation − mean, from the face and voice maps
persona_d	[5120]	its direction in the encoder's output space

The metadata holds the character, its clips, the windows and sigmas, the model, and how to apply it: add shift[w] to units idx[w] at every video token during step window w, in a prompt that does not name the character (the explorer's DiT experiment page, op "distilled character").

Use

hf download johndpope/h3-neuron-explorer --repo-type dataset --local-dir ~/h3-neuron-explorer-data

Then start the explorer with H3_VARIOUS_DIR=~/h3-neuron-explorer-data and copy (or link) maps/*.pt into the explorer's runs/h3/encoder_sweep/. The explorer README has the details.

Notes

• The clips show likenesses of real people and characters generated by MiniMax H3 (base model, no LoRA). For research into how the model represents identity; not for impersonation.
• The maps were built with H3's text encoder in ComfyUI; a unit's z for a character = (its activation − the mean over all characters) / the standard deviation over them. maps/ and the DiT files are activation statistics: no frames, latents or audio of the clips are in them.