metadata

license: creativeml-openrail-m

SD 1.5 dreambooth models fine-tuned on Anythingv3. Checkpoint files trained on (number of steps divided by 90) images scraped from boorus, with text encoder trained for around (number of images times 12) steps. Focused on ckpts before moving on to safetensors and lora. (~11/2022)

Textual inversion embedding versions can be found here (trained off the same or similar datasets).

nes2-1350 (Negev)

Fine-tuned 12/05/2022
Keyword: negebave
First attempt

ma1-850 (FMG9)

Fine-tuned ~12/2022
Keyword: magmgbg
First attempt

ca1-1260 (KAC-PDW)

Fine-tuned 12/06/2022
Keyword: kcpdweh
First attempt

p3-1800 (P90)

Fine-tuned ~12/2022
Keyword: hrstlleds
Third attempt

pn1-1350 (P90)

Fine-tuned ~12/2022
Keyword: hrstlleds
Fourth attempt, may get best results with TI

pr2 (Persicaria)

Fine-tuned 12/06/2022
Keyword: prsheheas
First attempt, may get best results with TI

nexl1_11142023_test

Keyword: negev (girls' frontline)
One of first attempts, works best with this model
Tested with ComfyUI, strength at ~1.5 seems to give best results (may be underfitted)

pxl1_11182023_test-000150

Keyword: p90 (girls' frontline)
Fine tuned for longer (cut off at ~150 epochs vs 50 epochs for previous), tested with this model
Tested with ComfyUI, may be overfitting to undesirable features (changing dataset for next test)

MTBA (previews, any future loras or models trained off better bases- hopefully more better SDXL too)