AbstractPhil/sd15-rectified-geometric-matching

V1 weights test push

https://github.com/AbstractEyes/sd15-flow-trainer

https://huggingface.co/AbstractPhil/sd15-rectified-geometric-matching/blob/main/colab_trainer.py

Step 0

Step 500

Step 1000

Step 1500

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

The geometric prior modulates CLIP encoder hidden states through 4-layer stacked k-simplex attention before they reach the 16 cross-attention blocks in the UNet.

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

Usage

from sd15_trainer_geo.pipeline import load_pipeline, load_geo_from_hub

# Load base SD1.5 + fresh geo prior
pipe = load_pipeline()

# Load trained geo weights from this repo
load_geo_from_hub(pipe, "AbstractPhil/sd15-rectified-geometric-matching")

# Or one-shot: load base + geo in one call
pipe = load_pipeline(geo_repo_id="AbstractPhil/sd15-rectified-geometric-matching")

Training Info

• dataset: AbstractPhil/imagenet-synthetic (flux_schnell_512)
• samples: 10000
• epochs: 1
• shift: 2.5
• base_lr: 0.0001
• min_snr_gamma: 5.0
• cfg_dropout: 0.1
• batch_size: 6
• loss_final: 0.3784324672818184

Post Analysis

License

MIT — AbstractPhil