In Waterfall SDXL

TL;DR

Photorealistic, allows subject interaction with moving water, mainly waterfalls, in SDXL derived checkpoints, except Pony/Illustrious.

Trained on BigLove XL4 (recommended).

Feel free to message me for tips and LORA support!

EDIT 21 Oct 2025: If you use an SDXL checkpoint that does not have the dmd2 LORA baked in (such as the various versions of Big Love, Lustify Endgame, Graam Juicy Merge, Analog Madness XL, Alma Erotic Photo LS, etc.), you can set the LoRA strength of the dmd2 LORA.

I found that setting the dmd2 speed LoRA strength at 1.1 or 1.2 produces immersion more often, even with the waterfall LORA strength at a relatively modest 0.5. With many LORAs, including this waterfall one, image quality is degraded at higher LoRA strength. By setting the strength of the dmd2 LORA at 1.1, you can get easier water immersion without too much quality degradation, and lower waterfall LORA strength. Give it a try, since it can be very difficult to get good immersion with dmd2 otherwise.

Comfirmed to also work with Lustify Endgame, Analog Madness and Cyberrealistic XL. Testing of other checkpoints is underway. Juggernaut XL Ragnarok works, but somewhat more reluctantly, compared to BigLove you may need to try higher LoRA strength or make more gens before you get water immersion; JibMix_v18 works - it is harder than Big Love to get water immersion at lower LoRA strength, but the water in JibMix looks excellent on some gens). GraAm Juicy Merge works almost as well as BL XL4. Lustify OLT works, and is comparable to Endgame.

If you find success with a checkpoint not mentioned here, please drop a comment.

The LORA was mainly trained on solo SFW (bikini) images, but will do nudes easily if you remember to add nude to the prompt and use a checkpoint that does nudes. At higher strength it is more likely to sometimes produce bikinis or swimming trunks, or hybrids. Explicit prompting of what you want to see avoids this. It will easily make bikinis or clothes if you prompt for them. I have not tried multiple subjects, but assume at low strength it should not be a problem if the checkpoint does it OK - as you can see from images below it is possible.

The dataset is not as Caucasian dominated as many checkpoints typically are, and also contained a smaller number of men.

Short prompts make the LORA way more flexible.

The less you describe the main subject, the more likely you are to get water interaction. If you like the magic image slot machine aspect, this will suit you. It is about as good (or bad) at more detailed prompts as the checkpoint is in general. Have patience. Drink coffee. Use "Variation Seed". Know when to admit defeat.

Training Tag Frequency List (the number after = represents how many times the tag was used in the dataset; not all tags have a strong effect, but you can play around and see which ones work for you):

in waterfall=104, droplets in air=100, 1girl=94, sunlight=71, clear water=50, nude=48. from front=47, from side=39, partially obscured by water=38, glittering=36, upper body=33, splashing on surface=32, long wet black hair=31, motion blur=28, looking at viewer=22, closed eyes=22, from behind=19, ass=19, froth=17, full body=17, smile=17, glassy water=16, partially submerged=16, splashing on shoulder=16, dappled sunlight=16, standing=16, water curtain=15, nipples=15, head tilt=14, close-up=14, moss=14, from above=14, sitting=13, weak sunlight=13, low angle=13, barefoot=13, overhang=13, long wet brown hair=12, sideboob=12, plant=12, collarbone=12, teeth=12, splashing on rock=12, small breasts=11, splashing on head=11, splashing on ground=11, tree=11, makeup=10, arm support=10, closed mouth=10, blurry background=10, shadow=10, large breasts=10, long wet blonde hair=9, thighs=9, brown eyes=9, grass=9, necklace=9, 1boy=9, forest=9, navel=9, pussy=8, waves=8, parted lips=8, open mouth=8, peach bikini=8, fern=8, wet hair=8, bracelet=7, stream=7, golden hour=7, brown rock=7, looking back=6, arched back=6, wet black hair=6, splashing on arm=6, wide angle=6, roots=6, arms up=6, turquoise water=6, ribs=5, kneeling=5, looking to side=5, light brown rock=5, facial hair=5, white water=5, cavern=5, medium breasts=5, showering=5, looking down=5, back=5, armpits=5, wet=5, earrings=5, splashing on face=5, arms at side=5, spreading=4, eyeliner=4, bent over=4, crevices=4, brown skin=4, leg up=4, dark nipples=4, flowing on body=4, partially immersed=4, touching rock=4, brown hair=4, anus=3, tanlines=3, lips=3, profile=3, dripping=3, long hair=3, nose=3, behind waterfall=3, subject backlit=3, free-flowing waterfall=3, brown swimsuit=3, sitting on rock=3, splashing on buttocks=3, ferns=3, head out of frame=3, fallen log=3

Settings:

Main tag: in waterfall (adds a waterfall in the picture; leave out if you want water interaction in other scenarios - can make sense for showers etc. and make for some surreal fun in environments not normally associated with water.)

Strength: 0.5-0.7 (higher reduces picture quality and introduces distortions and more ugly/repetitive textures; lower causes weak or no effect)

Close-ups: Add close-up to the prompt. Use the dmd2 LORA with Sampler: LCM, Scheduler: Exponential, Steps: 6-10, CFG: 1.0-1.2

Further away: Hard mode (SDXL limitation). Use default settings for your SDXL checkpoint.

Step 1:

For BigLove XL: DPM++ SDE, Karras (XL), 20-30 steps, cfg 3-10, clip skip 2, 832x1216 or 1024x1496. For faster generation 2M SDE/3M SDE works too. Lower CFG recommended, 4 is a good starting point.

For Lustify: Sampler: DPM++ 2M SDE/DPM++ 3M SDE, Scheduler: Exponential/Karras, Steps: 30, Cfg: 2.5-4.5, Highres.fix: upscale by 1.4-1.5, denoising ~0.4.
Other checkpoints: Check the checkpoint info page.

Step 2 (optional, may improve realism sometimes, usually reduces texture detail):

Bring img from step 1 to img2img, where you use the dmd2 LORA with Sampler: LCM, Scheduler: Exponential, Steps: 6-10, CFG 1.0-1.2, Denoise 0.15-0.4. Upscaling factor depends on what your hardware can handle, and how high you can go without introducing distortions (lower Denoise=less change=less risk of distortion).

IMPORTANT: The dmd2 LORA actively fights the concept of immersion. It is nearly impossible to get moving water immersion with dmd2, except in close-ups.

Still here? Congratulations, your attention span is highly impressive!

It was a dark and stormy night...

This is my first LORA. I know I should have started with something simpler. But this is what I wanted, and while far from perfect, it kind of works now. Will try to improve it in future versions.

---

To see some amazing photorealistic nsfw and sfw gens, share your own, or discuss tips and tricks for photorealistic gens, pop by this Discord Server

---

A huge thanks goes out to SubtleShader, without whose helpful coaching and encouragement this would never have happened. He created the BigLove checkpoints and a whole host of LoRAs for SDXL - check them out if you have not already.

Thanks are also due to the helpful and experienced regulars at the (strictly SFW) OneTrainer Discord server.

If you find this Dora LORA useful, feel free to tip me some Buzz so I can make it better, and maybe create something else you have dreamed of, such as a yak milking LoRA. ;-)