This is a Flux dev version of my SD 1.5 and SDXL LoRA. Flux is finally able to do large low-cut armholes on its own, but it does require some heavy prompting sometimes. This LoRa will easy your generations…<h3 id="are-you-tired-of-sd-flux-giving-you-this-completely-prude-tight-fit-high-cut-armholes-tank-tops-do-you-want-your-generated-dork-himbos-and-bimbos-showing-off-their-muscles-at-the-gym-u3tfq3y8e">Are you tired of <s>SD</s> Flux giving you this completely prude tight fit high-cut armholes tank tops? Do you want your generated dork himbos and bimbos showing off their muscles at the gym?</h3><h3 id="let's-fix-it!-u08xxdndm">Let’s fix it!</h3> Lora strengh works from 0.1 up to 2.0.Start with 1.0 There is no trigger word. But you should reference a tank top, sleeveless shirt, gym stringer or singlet. A simple "tank shirt" is normally enough. If your males are getting too much belly crop top shirts or shirtless, it helps to use just “tank” without the word “top”. Supporting prompts:<pre><code>shirt, tank top, tank, sleeveless shirt, high neck crewneck raised collar, low-cut chest (males), low-cut cleavage (female), extreme low-cut dropped side holes, side boobs (female), side pecs (males), (male/female) nipple slip, side view, profile, front view, full body medium shot, long shirt, loose shirt, muscle body, thin body </code></pre> Negative supporting prompts (Use thresholding or skimmed):<pre><code>skin holes, shirtless, jewelry, necklace, muscular, head out of frame, 
WHEN YOU WANT A HIGH NECKLINE SHIRT: low-cut chest, nipples, cleavage, showing his/her chest/cleavage</code></pre> I probably need to do better captions because unfortunately this Flux LoRa is worse at differentiating high crewneck shirts and low-cut chest/cleave shirts than the previous SD15 and SDXL versions. It will default to large cleavages most of the time as you increase the strength... Prompt a little bit harder like "high neck crewneck raised collar" and it will work. Using negatives can help a lot as well. And if all fails, a low lora strength.If you actually write something like "wearing a loose stringer tank top shirt with large low-cut dropped armholes showing his side pecs [or her side boobs]" Flux will most of the times give you what you want. But with my lora you can simply use "wearing a tank shirt" and control it with the strength. With my LoRa you can do extremes and also experiment giving other types of clothes large armholes like dresses, coats, hoodies etc. Those might be harder with just prompt... Anyway. Have fun.‎This version had a block analysis and <a rel="ugc" href="https://civarchive.com/models/1048175/flux-lora-block-weights-preset-tester-comfyui">was remerged with my preset 7</a>. It is good at keeping the original content, background, composition and quality. So keep that in mind, this LoRa is not a photo quality enhancer. Use it with other LoRas.‎Problems with the current Lora:<ul><li>I'm still not fully satisfied with it. The effect of the low-cut could be better controlled. On the extremes, it tends to do more cleavage than large side armholes.</li><li>High crew-necks needs to use more collar prompting or negatives. This could be improved with a better captioned dataset.</li><li>Nipples sucks because nipples sucks on Flux dev. This won't fix nipples, use another lora to increase its quality.</li></ul>Some training information: Trained on Dev2Pro (don't use it for inference). full fp16 model. Full blocks and the remerged. Batch of 2. Prodigy. 18 out of 22 epochs. 1024 res. More than 205 pictures. 1/1/5 folder repetition. 1_stringer shirt / 1_low-cut chest stringer shirt / 5_low-cut cleavage stringer shirt. rank 8/8, no regularization images” Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.

Stringer Gym Tank Top / Dropped Side Holes / Large Armholes – Flux Dev.1

v1 use strength from 0.2 to 2. Recommended 1.0No trigger word but needs a mention of a tank shirt, stringer, sleeveless, singlet etc Epoch 18 of 24steps 3420 of 4560 (batch 2) = 6840212 images. 8h training

Files with SHA256: 7f680dcb409931a0d23acd590c54e492f7ad596802f6c50a7023a48ec66e582f