Breast Implants

Feedback welcome! Support is appreciated, but never expected: About 4.0 - Improved stability and delineation between sizes. - Pruned dataset and using all new training settings. - This is also a LoKr instead of a LoRA now. - Should be better in pretty much all situations to previous versions. - Used with 4-step and base with turbo LoRA with good results. (Images I posted were with 4step Klein 9B, not base with turbo LoRA) - Prompt adherence is better with more flexibility. - Used euler with sgm uniform or beta for all testing - Strength: 1.0 (I've used up to 1.5 but likely to get more images with bad anatomy) - Read "About Version 3.0" below for more info regarding settings and prompting tips. About Version 3.0 - New dataset should allow for much better results across the board. - No trigger word like previous versions, so use natural language - If you want a certain, look describe it ad nauseum and the image usually responds well. (larger, perkier, etc.) - Useful terms: - Size: small, medium, large, huge, gigantic, massive, etc. - Shape: round, perky, "bolted on" (these aren't as good as I wanted them to be) - General Modifiers: very, extremely, significant, prominent, etc. - Other: busty, bust, bustline, voluptuous, curvy, etc. - Dataset captions follow this general layout: A medium shot of a girl with [size] breast(s) (implants) facing to the right. She is wearing [a tight, thin shirt] and [smiling]. Her breasts are [very perky] and [round]. The outline of her nipples peaks are visible.... [rest of prompt] - Some new features that sometimes work when prompted: - Covering nipples with one hand/both hands - Middle buttons unbuttons revealing cleavage - Outlines of nipples peaks under clothing - Better side profile images - Better nipple quality and sizes - Better areola quality and color Settings: - Sampler: Tested with euler and res2m (other likely work well too). - Scheduler: simple, SGM uniform (most stable), beta, beta57 - Strength: I think the general use sweet spot is (1.15)-(1.25) - Safe 1.00 - General use: 1.25 - Bigger but more glitchy: 1.50+ Feedback welcome! Support is appreciated, but never expected: WARNING: If you are having issues with the clothing not covering the breasts then try the following: - Turn down the strength. You can achieve very large sizes with prompting alone, albeit it's inconsistent. - Do not use the word "breast(s)" as this will trigger more skin to show. - Use terms like bust, busty, bustline, etc. (example phrases): "A very busty girl..." "girl with a large bust" "she has a massive bust" - Describe clothing as large or well-fitted (examples phrases): "form-fitting dress" "large tube top" "large tight t-shirt with a crewneck" - Briefly tested with a breast slider and it seems to keep the clothes on better while still keeping the "fake" look About v2.5: This version (v2.5) was supposed to be v3.0, but it's honestly not consistent or refined enough. After multiple training attempts, I decided to release v2.5 while I try to improve the LoRA for a proper v3.0. It should be more flexible with sizes, angles, and posing as well as how clothing works with breasts. Also, breast shape should be more flexible too. Usage: - I do most all of my testing with simple workflows so I can know the images I get will be similar to others. - Primary used euler/simple or euler/sgm uniform at 4-8 steps - The RES4LYF samplers work well - Some of the DPMPP samplers work well. - Let me know if you find something else... - Strength of 0.5-1.5 seems to do ok. you can push it to 1.75 or even 2.0 at times. - I will admit this version will expose breasts more readily at larger sizes/strengths (1.00+) - Use natural language. Images were captioned with phrases like: -Medium shot of a girl with large breasts wearing... -Close up selfie of a girl with huge breast implants. -Wide shot of a girl facing away. She has very large breasts and is wearing... - Be wary of samplers/scheduler combos and step count. A distilled 4-step model like Klein is very sensitive to these parameters so adjust them accordingly. - Fun to use with a breast slider! About V1 and V2: Used 70-80 photographs depicting women with breast implants. This LoRA can do both clothed and topless women. If topless, mention nipples or areolae. This LoRA was not trained on genitalia so add your preferred LoRA for that. It seems to play decently well with other LoRAs and does not appear to mess with faces all that much if at all. There is no trigger word as all of the photos were captioned using natural language, but mention breast implants somewhere in the prompt. That being said: Training images were subdivided into "small", "medium", "large", and "huge" breast implants. Using the terms has a much more minimal effect that anticipated, although some size change can be seen. v2.0 changes - Now has a range of 1.0-2.0 although 1.0-1.5 is the sweet spot. - Improved size differentiation. - Some poses will break anatomy v1.0 changes - Use a strength of 1.0 with the term "small" or "medium" breast implants to achieve the smaller size. - Use a strength of 1.5 with the term "large" or "huge" breast implants to achieve the larger size. - Adding "very" before the term can further influence size - If you are wanting more control you can use a breast slider lora for smaller or much much larger sizes. I use it with the distilled model. Haven't tried with the base. CFG:1 Steps: 4 Sampler/scheduler combos I have found success with: - euler/beta - res_2s/beta - euler cfg pp/beta57 - whatever you like best.... (I'm sure there are better sampler/scheduler combos out there, so let me know if you have suggestions) Feedback welcome.