This LoRA reproduces hand over mouth scenes for Flux Dev.<h2 id="prompting-tips-for-v3v2.5:-kjz8qwdkl">Prompting tips for v3/v2.5:</h2>All dataset captions follow this structure. Do your prompt like that to ensure the best results:[An amateur/A professional] [close-up/wide-shot] [photograph] featuring a [victim's gender] and a [assaulter's gender] behind [him/her]. A [description of the assailant] covers the [victim's gender]'s mouth with [his/her] hand. [His/Her][description of the hand] hand [other description if needed] [tightly/gently] covers the mouth of a [brief victim description] [victim's gender]. [Describe here what the assaulter is doing with his free arm]. The victim has [victim's detailed description]. [Describe here where the assailant and victim are looking]. [Environment and setting description].Good news: The V3 also works with more random prompts.<h2 id="tips:-7yqo4clhe">Tips:</h2>It's not recommended to use a fine-tune as the hands are often messed up. This LoRA works pretty well in combination with <a rel="ugc" href="https://civarchive.com/models/1551668/samsungcam-ultrareal">SamsungCam UltraReal</a>.Sampler/Schedulers: <ul><li>Euler A/Beta for native picture</li><li>Upscale or second pass with Deis/Beta (6 steps, 0.15 to 0.3 of denoise works great to add skin details and sharpness).</li></ul>To avoid classic Flux's grid/stripe patterns on high resolutions, it's better to do a tiled Upscale/tiled second pass.<h3 id="other-versions-are-deprecated.-gvi7kmmw1">Other versions are deprecated.</h3>Please use this LoRA responsibly, following ethical and moral values. I take no responsibility for its misuse.

Handgag FLUX

<ul><li>Very high success rate!</li><li>More varied poses</li><li>No overfit</li></ul>

Handgag_FLUX_V3.safetensors

v2.5 Realistic

v1.0 Reverse grip

v1.0 Anime

v2.0 Realistic

V1.0 Realistic

This LoRA reproduces decent hand over mouth scenes.The model currently has some limitations:<ul><li>There is only one pose</li><li>Portrait aspect ratio does not work well</li><li>The images seem to lack in brightness</li><li>The success rate is approx. 75% with fp16/fp8/Q8, drastically lower if &lt; Q6</li></ul>Prompting tips:[Description of the woman covering the mouth] covering another woman's mouth with her hand. Her hand [description of the hand] covers the mouth of a woman [description of the second woman]. [Description of the environment and other details]Example: A Middle Eastern woman with curly dark brown hair covering another woman's mouth with her hand. Her slender hand has painted red nails and gently covers the mouth of a shocked caucasian woman with blonde hair. They are in a cozy living room during the evening with a warm fireplace visible in the background through the room's windows.It's working also with two men, woman on man or man on woman.Observations: Weirdly enough, the Anime version of the LoRA can also produce very good realistic images, but it was trained only on anime pictures. However, the realistic version of the LoRA (trained only with real pics) struggles with anime style.

Files with SHA256: eda5a4417448ab2f34e131bf9e9944a0435576a2cf2e2a2f319ca209c4d98e8c