HootiTally - Krea 2 character LoRA

File: hootiTally_krea2.safetensors (in ComfyUI\models\loras) · Base: Krea 2 RAW · Rank: 32 · Trigger: hootiTally (optional, see below)

Best strength: 1.0 (0.9–1.1) · Tested on FinePorn V5 BF16 (a Krea 2 fine-tune)

The character:

hootiTally is a fictional adult character. The look the LoRA locks in: a slim young woman with fair skin, long straight dark hair with full bangs, and dark-rimmed glasses, with a narrow face. The training set is one look only, so the LoRA does not reproduce any of the character's other hair colours.

Naturally trans-leaning. The character in the training photos is a trans woman, and the LoRA learned that as part of her body. Prompts that only say "woman" still produce it whenever she is undressed or wearing something thin or sheer, because only one of the 20 training captions mentions anatomy, so the LoRA treats it as part of the character. Describing full clothing keeps images SFW.

Training:

| Trainer | Ostris AI Toolkit on Krea 2 RAW |

| Photos | 20, one look (black hair, glasses, bangs), hand-picked from 30 matches for visible face and different outfits and poses |

| Captions | Prose, starting with hootiTally,; sentences about her glasses, hair, eyes and face were removed so the trigger carries them |

| Steps | 1,800 (90 passes over each photo, the most common range in the RLY Krea 2 LoRAs) |

| Network | LoRA, rank 32, alpha 32 |

| Learning rate | 2e-4, decaying linearly to 0 (start factor 1.0, end factor 0.0) |

| Optimizer | AdamW8bit, weight decay 1e-4, batch size 1, bf16 |

| Resolution | 768 |

| GPU | RTX 3060 12 GB: 8-bit base weights with layers streamed from system RAM, about 10 hours |

| Result | Finished with no loss is nan; loss fell from about 0.108 to about 0.095 |

Testing:

(FinePorn V5 BF16, 12 steps, cfg 1, same prompts and seeds each time)

| Strength | Result |

| Off | A different random woman each time |

| 0.8 | Glasses, dark hair and bangs in 4 of 6 images; clothes and rooms obeyed |

| 1.0 | Best. Glasses, straight dark hair and bangs in all 6 images, a consistent face, prompts still obeyed |

| 1.3 | Strong identity, but the training backdrops take over and clothing instructions start to be ignored |

| 1.6 | Over-baked: most images ignore the clothing asked for, backdrops are all curtain and tapestry, faces drift |

Other findings:

- Without the trigger word at 1.0 it still works (glasses, hair and bangs in 5 of 6 images), and the trigger-word logo (below) does not appear.

- Checkpoints from step 1500 on are almost identical to the final one; step 1200 was slightly less consistent.

How to use it:

1. Load hootiTally_krea2.safetensors at 1.0.

2. Write prompts in caption style ("Photograph of a young woman standing in a bedroom, wearing…") and name the clothing and the setting.

3. Leave the trigger word out unless the face is weak. If you use it, put it first.

4. Don't describe a different hair colour, eye colour, skin tone or ethnicity: the prompt text overrides the LoRA.

Known issues:

- Trigger printed as text: with hootiTally in the prompt, the model sometimes draws it as a small "hootitally" logo, for example on a jacket chest. Leaving the trigger out removes it.

- Backdrop leakage: most of the 20 photos share one ivy-curtain room, so an ivy curtain or tapestry shows up in bedroom prompts at every step, and more strongly at 1.3 and above.

- Nudity leakage: the training photos are adult and include nudity. At 1.3 and above the LoRA starts to override prompts that ask for clothing, and even at 1.0 vague clothing ("an oversized shirt", "a clear raincoat") can come out undressed. For SFW images, spell the clothing out head to toe, for example "a gray t-shirt, sweatpants and socks".

- Only tested on FinePorn V5 BF16. It was trained on Krea 2 RAW, but I haven't tested it on RAW or Turbo directly.