Sign In

OnlySaree_PonyDiffusion_by_Sarcastic_Tofu

12
164
1.0k
8
Verified:
SafeTensor
Type
LoRA
Stats
164
1,003
Reviews
Published
Jun 28, 2024
Base Model
Pony
Training
Steps: 5,000
Epochs: 10
Usage Tips
Clip Skip: 1
Strength: 0.5
Trigger Words
Woman in Saree, OnlySaree_Pony_Style, Wearing a BlouseLess BraLess SleeveLess OneShoulder Saree, Exposed Belly Button
Hash
AutoV2
06E0D99391
Created on Civitai

A Saree is a single piece of traditional South Asian cloth that can be worn with or without a Blouse and Petticoat. This model was created to help generate realistic images of South Asian Women or Women of any other ethnicity across the globe wearing blouseless Sarees, wearing only Sarees without any other additional supporting attires. I initially trained this model for SD 1.5 using some carefully curated, AI enhanced, color corrected, logo and watermark free images of sensual women wearing just Sarees without any other additional supporting attires with the help of Dreamlook AI. I eventually further trained this dataset for both SD 1.5 and SDXL 1.0.

You can generate good quality images with following Stable Diffusion prompt inputs with other proper keywords. phrases, negatives, textual inversions and LORAs on this Pony Diffusion based models -

Woman in Saree, OnlySaree_Pony_Style, Wearing a BlouseLess BraLess SleeveLess OneShoulder Saree, Exposed Belly Button

(

Versions Update -

OnlySaree_V1.0 - First version of the OnlySaree LORA, works better in standard Square or Portrait modes but can do some Landscape mode outputs.

OnlySaree_V2.0 - This second version uses some additional training data to improve the quality of outputs, special focus has been given to focus on more accurate Landscape mode outputs, they aren't perfect but a lot better than that of version 1.0

OnlySaree_V3.0 - An updated version of my OnlySaree LORA for SD 1.5. This one has more training data and should produce better Non-South Asian female subjects wearing only Saree in output images.

OnlySaree_V4.0 - This version has more additional data sets and it's training process parameter had 10,000 steps. This is aimed to gain more result accuracy.

OnlySaree_V5.0 - This version has more refined data sets and it's training process parameter had 10,000 steps. This is a good overhaul of the original model and can use more concepts (Wet Saree, Sideboob etc.) via new optional keywords. (Please see the keyword file too) This is aimed to gain more result accuracy. It's available here - https://civitai.com/models/277659

OnlySaree_CIVITAI_1.0 - an alternate version of this LORA using CIVITAI's own training tool, it's available here - https://civitai.com/models/287393?modelVersionId=323237

OnlySaree_SDXL_1.0 - This is the first SDXL adaptation of my OnlySaree LORA for Stable Diffusion 1.5. Expect the improvements that SDXL brings to you with unique touch of OnlySaree style with this LORA.

OnlySaree_SDXL_2.0 - An updated SDXL adaptation of my OnlySaree LORA. This one has more training data and should produce better Non-South Asian female subjects wearing only Saree in output images.

OnlySaree_SDXL_v3.0 - This was trained on a bigger dataset than the previous version. This improves the overall quality of image generation. It is available here - https://civitai.com/models/295929/onlysareesdxlbysarcastictofu

OnlySaree_Pony_1.0 - This is the first Pony Diffusion adaptation of this model.

Both LORA models have their own hits and misses. This initial LORA version provides better anatomically correct models while the CIVITAI LORA version provides more accurate Saree aesthetics.

While the data set had no NSFW content, some base models do produce more NSFW results when paired with any of these LORAs. Some base models can do better job than others on Landscape Mode Image generation, for SD 1.5 Landscape mode generation is tricky and often results subjects with deformed anatomy or simply extra body parts, SD 1.5 does not seem to respond to negatives while dealing this so switch things around to get your desired results. SDXL 1.0 based model seems to do better anatomically correct subject but you still have to do some good prompt crafting and resource selection.

)