Spaces:
Running on Zero
Download README.md from AbstractPhil/agatha-diffusion-playground: direct link, hf CLI and curl.
- Browser
- Download file 3.57 kB
-
https://huggingface.co/spaces/AbstractPhil/agatha-diffusion-playground/resolve/main/README.md
- Command line
-
hf download hf://spaces/AbstractPhil/agatha-diffusion-playground/README.md
-
curl -L -o README.md https://huggingface.co/spaces/AbstractPhil/agatha-diffusion-playground/resolve/main/README.md
title: Agatha Diffusion Playground
emoji: 🐢
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 6.32.0
python_version: '3.12'
app_file: app.py
pinned: false
license: cc-by-nc-sa-4.0
short_description: Agatha against Lia, her control, blind or by name.
models:
- AbstractPhil/geofractal-agatha-diffusion-training
- black-forest-labs/FLUX.2-klein-4B
- google-t5/t5-base
Agatha diffusion playground
Two small text-to-image models, trained the same way on COCO 2014 at 256 px, shown side by side from the same noise:
- Agatha (geofractal
agatha_v2, about 42.0M parameters): a one-block transformer trunk, six towers, a damped oscillator, and learned weights per noise band. - Lia (about 42.9M parameters), the control: a plain transformer with the same embeddings and blocks, 13 of them, and no
towers and no oscillator; the known working design at the same size and recipe. Her design follows DiT and U-ViT-S/2, and
her name is Leah of Dante's Purgatorio XXVII (lines 97-108), who gathers flowers into a garland with her own hands: the
active life. Lia names the line of control models for these diffusion engines. Her training runs keep the word
plainin their names.
Where each model comes from is told on the weights card, with the results so far: at 100,000 steps Lia leads Agatha on every read, on both seeds.
Write a caption, choose two checkpoints, and generate. Each picker lists every checkpoint the training runs have published (either model, any step, raw or averaged weights) plus None; by default they hold each model's newest raw checkpoint, and None where a model has nothing published yet, so one checkpoint can also be viewed alone. Both sides start from the same noise for every picture. In blind mode the two sides are shown in random order: pick the better side, then reveal which was which. Your picks are counted for this session only.
Tower switches. For Agatha checkpoints each side has six switches, one per tower: cantor, beatrix, simplex and helix (attention towers, each with its own rotary position code) and wide_resnet and frequency (convolution towers; frequency sees a high-passed input). Unticking a tower sets its push to zero for that side's pictures; the trunk and the other towers stay as trained. Put the same checkpoint on both sides to compare tower settings, blind or by name. Lia has no towers, so her switches change nothing.
Weights. At early checkpoints the raw weights give the sharper pictures. The averaged (EMA) weights lag behind early in training and catch up later.
Where things come from. The weights are in
AbstractPhil/geofractal-agatha-diffusion-training,
folder coco256/; the list refreshes when the page loads, so new milestones appear without a restart. The code is
github.com/AbstractEyes/geofractal, src/geofractal/router/prefab/agatha_v2,
installed at a pinned commit. Pictures are decoded by the FLUX.2 autoencoder from black-forest-labs/FLUX.2-klein-4B
(Apache-2.0), and captions are read by T5-base, both at the revisions the training data was encoded with.
License. The models were trained on COCO photos, most of which carry non-commercial licenses, so the weights and the pictures made here are for non-commercial use only.
These are small research models early in training: expect rough pictures, not a finished image generator.