# Model description

A style LoRA for MiniMax H3 that imitates the visual language of TheOfficialPit: the artist's trademark anatomy.

The training set covers 482 captioned illustrations (detailed prose), split into two stages:

1. Structure stage — 16,000 steps at rank 32, with timesteps weighted to the noisier half of the schedule (this is where composition, proportions and silhouettes get locked in).

2. Low-noise refinement stage — a continuation of stage 1, adding 6,000 steps that draw exclusively from the low-sigma band (sigma ≈ 0.4 → 0) at a lower learning rate. Lip, cheek and skin detail plus edge crispness come from this stage.

Side-by-side comparisons against a straight doubled-length run and a rank-64 variant (same dataset) showed the two-stage build came closest to the reference images for faces and body shape.

## How to use

- Base model: MiniMax H3 (fl2va). Verified against the Comfy-Org repack minimax_h3_fl2va_pruned_int8_convrot, Qwen3-VL text encoder, video VAE) as well as ComfyUI's built-in H3 nodes (v0.33+, zero custom nodes required). For stills, set length 5 — the smallest frame grid H3 accepts — and use frame 1.

- Strength governs the look: 0.8–1.0 gives the full treatment; 0.4–0.6 keeps things subtle. The training captions and prompts contain the trigger 7h30ffici4lpi7, yet the style engages as long as the LoRA is attached — so pull back by dropping strength, not by deleting the trigger.

- Skip the negative prompt: H3 is guidance-distilled and runs at CFG 1, which leaves negatives doing nothing. Do all steering in the positive prompt.

- Prompting: plain prose performs well — cover subject, pose, clothing, background, plus any rendering keywords in the style of the showcase examples ("bold uneven linework, cel shading, saturated palette...").

- Mature content appears easily with this model; the upload carries the matching flag.

---

v1.0 — the refinement-stage checkpoint of the two-stage run (structure + low-noise pass), chosen after head-to-head comparisons against longer and higher-rank candidates. Rank/alpha 32, bf16; trained via ai-toolkit 0.12.25 on the fl2va partition using DOP regularization.