Generates ankle socks reliably. Current models apparently only generate calf-, knee- or even thigh-high socks when prompted with "ankle socks" or "ankle_socks."
[Note: check per-model details in the sidebar for activation keywords.] T
These versions use the same core dataset as v1.0 (in various resolutions), but with attention masking to improve the training process and subsequently, model flexibility. All models were trained in OneTrainer; targeting 1600-2000 training steps spread over 10ish epochs. Later models (Krea2, Illustrious versions) use hard reset sine LR.
LoRa weights were normalized at 1.0 for the SD1.5 versions (the recommended weight to use this model) using the invaluable Stable Diffusion ai-toolkit (https://github.com/ostris/ai-toolkit). The rest did not require normalization.
Sample images include all metadata for recreation in your own environments; highres. fix and ADetailer (just to make faces pretty) were used to generate the samples. No other embeddings, LoRa or manual inpainting were used during/after generation.
You're free to use this however you wish. It's socks. Would love to see some creations using this LoRa!
Description
Gentle curation of dataset + re-masking to better capture the sock/ankle area/proportions that most models struggle with. Added an additional trigger word ("4nk1es0x"), but it's weak; I suspect my tag dropout settings impacted it; next version I will likely disable that training behavior as a test.
See comparisons grids of v1.1 vs. v1.0 vs. no-lora here



















