Type | |
Stats | 520 3,022 423 |
Reviews | (75) |
Published | Jun 26, 2024 |
Base Model | |
Training | Epochs: 10 |
Usage Tips | Clip Skip: 2 Strength: 0.8 |
Trigger Words | courtyard pool rooftop |
Hash | AutoV2 2A1963757E |
This is a photographic lora for Pony v6 XL. It was trained on 20 photos, mostly Morocco luxury riad courtyards with pools, and a few rooftop terraces and interiors.
Works well with photorealistic Pony models, like Pony Realism.
Lora strength: 0.5 to 1.0 (depending on the model)
Tags
Tags: courtyard, pool, rooftop
Also try these tags that were in the captions: tile mosaic, tile floor, fountain, garden, potted plant, citrus tree, palm tree, vine, table, chair, lounge chair, couch, lantern, hanging light, candle, pillar, railing, indoors, outdoors, day, night, sunlight
Prompting "bedroom" or "bathroom" might also work.
Why?
This is an experiment in making a specific setting with a small dataset.
What's wrong with it?
Complex architecture, furniture, and mosaics will not be coherent all the time. Maybe I trained the lora badly, or maybe these things are just hard to draw for Pony-based models. I'm just a beginner at this, so please let me know what I do to improve it.
The building will probably always have white walls, since I only used pictures with white walls for consistency.
Most pictures in the dataset did not have people, so you may have to put emphasis on your character in the prompt to make sure they show up. Or try turning down the lora strength. The lora also tends to make wide shots. It may have problems with the scale of humans in swimming pools (because it was trained on very small pools.)
Training info
20 images x 15 repeats= 300, batch size 5, 10 epochs (3000 total image looks) 0.0005 lr, 0.00005 text lr, prodigy cosine, noise offset=0.05, min SNR gamma=5, dim 16
(Trained on Civitai. I don't have a good video card.)