<a target="_blank" rel="ugc" href="https://www.seaart.ai/User/fcafabca072c11f392e9ea006fffeebb">https://www.seaart.ai/User/fcafabca072c11f392e9ea006fffeebb</a>DISCLAIMER: The input-image used for these examples is a FLUX generated image with one of my character LoRas and does not depict a real person.<h2 id="model-description">Model description</h2>The LoRa was trained on 80 hand-picked NSFW on/off photo pairs gathered mostly from reddit. Every pair was captioned by hand to make sure to differentiate between the two major categories:<ul><li>Identical Pairs: Pairs where the start and end image are exactly (or very close to) the same except for clothing of course.</li><li>Different Images: Pairs where changes in poses etc. between start and end image differ enough that it could cause the model to learn that its fine to not stick closely to the reference image even if prompted for it.</li></ul><h1 id="captions">Captions</h1>These are example captions from the dataset for both an Identical Pair and "Different Image":Identical: "Completely remove the subjects clothes. She is now standing completely naked, her small breasts and vagina with pubic hair fully exposed. She has a small belly-button piercing. Maintain identical subject placement, camera angle, framing, and perspective. The rest of the image remains the same, only completely remove her clothes to expose her naked body."Different: "The image now looks like a different photo of the same subject. All of her clothes are now removed. She is now posing completely naked. Her medium sized breasts and clean shaven vagina on full display. She is now posing with her hands tangling in her own hair. Maintain identical subject placement, camera angle, framing, and perspective. The rest of the image remains the same, only remove her clothes to expose her breasts and vagina and change her pose."As you can tell, the block {Maintain identical subject placement, camera angle, framing, and perspective. The rest of the image remains the same, only ...} comes up in both caption types to reinforce the prompt so I recommend using this or something similar in your own prompts. Because of these prompting categories the model works very well both when simply removing clothes and creating a new image of the same subject in a different pose (see example images).<h1 id="prompts">Prompts</h1>These are the example prompts for static and different images that are general enough to have worked well on different images when generating with the model. Make sure to adjust these to your needs, depending on your image, for example if you use a portrait shot you may not want to include the vagina parts of the prompts.Identical: "Completely remove the subjects clothes. Her clothes are now completely removed to expose her small breasts and clean shaven vagina. Maintain identical camera angle, framing, and perspective and make sure to leave her facial features identical. The rest of the image remains the same, only remove her clothes to expose her naked body."Different: "The image now looks like a different photo of the same subject. Her clothes are now completely removed to expose her small breasts and clean shaven vagina and she is now standing in a seductive pose. Maintain identical camera angle, framing, and perspective. The rest of the image remains the same, make sure to leave her facial features identical, only remove her clothes to expose her breasts and vagina and change her pose."For better results I recommend a vertical aspect ratio like 9:16 or 3:4. To make your prompt work more consistently on a new image specifically reference the piece of clothing to remove. For example 'Completely remove the subjects clothes' turns into 'Completely remove the subjects white crop-top and jean shorts'. Reinforce this in the 'The rest of the image...' block to further increase consistency, e.g.: 'The rest of the image remains the same, only remove her white crop-top and jean shorts to expose her naked body.'<h1 id="basic-workflow">Basic Workflow</h1>This is the comfyui workflow I used to run Kontext with the LoRa (Full credit to Ostris/AIToolkit, its from his LoRa training guide):<a target="_blank" rel="ugc" href="https://pastebin.com/vYnttdEF">https://pastebin.com/vYnttdEF</a>

Files with SHA256: 2c1e3abbc562cf125bc9a6bb531ace43b93517f69e3d4b49460ef10e86cfb357