AbstractPhil/sd15-geovocab-lora-prototype

KSimplex Geometric Attention Prior

Geometric cross-attention prior for SD1.5 using pentachoron (4-simplex) structures.

Architecture

Component	Params
SD1.5 UNet (frozen)	859,520,964
Geo prior (trained)	4,845,725
Geo conditioner (trained)	1,613,847

Simplex Configuration

Parameter	Value
k (simplex dim)	4
Embedding dim	32
Feature dim	768
Stacked layers	4
Attention heads	8
Base deformation	0.25
Residual blend	learnable
Timestep conditioned	True

GeoVocab Conditioning

Parameter	Value
Gate dim	17
Patch feat dim	256
Num patches	64
Cross-attention	enabled
Cross-attn heads	8
Blend mode	learnable

Usage

from sd15_trainer_geo.pipeline import load_pipeline

pipe = load_pipeline(geo_repo_id="AbstractPhil/sd15-geovocab-lora-prototype")

Training Info

• dataset: AbstractPhil/synthetic-characters (schnell_full_1_512)
• subdir: schnell_full_1_5e-5
• samples: 50000
• epochs: 10
• steps: 83330
• shift: 2.0
• base_lr: 5e-05
• min_snr_gamma: 5.0
• cfg_dropout: 0.1
• batch_size: 6
• geo_loss_weight: 0.01
• geovocab_lr_mult: 2.0
• clip_vae: AbstractPhil/geovae-proto/clip_vae/best_model.pt
• patch_maker: AbstractPhil/geovocab-patch-maker
• loss_final: 0.3035515168607235

License

MIT — AbstractPhil