A South Indian woman in her thirties, built as a character, not a look-alike. Fictional: she does not exist and no real person's images were used. Trained for continuous storytelling — the same face at every framing, in sarees and at home, in Kerala light.
There is no other standalone South Indian character LoRA on this site. That is why she was made.
USE
Trigger: zxrevathi. Always add her age — the model reads it:
zxrevathi woman, 33 years old, <framing>, <expression>, <hair>, <outfit>, <light>, <place>
• Base: Krea 2 Turbo (trained on Krea 2 Raw, run on Turbo — the usual Krea 2 recipe)
• LoRA strength: 1.0–1.2 (1.2 is the best measured; below 0.8 she fades)
• Steps / CFG: 8 steps, CFG 1, Euler / Simple
• Size: 1024x1344 or 768x1024 native; upscale after
• Rank 32 / alpha 32, 218 MB
Candid lines like "raw phone photo, … natural unpolished skin texture, slight grain" give the best skin. Editorial lines carry their own camera words.
MEASURED, NOT CLAIMED
Identity is scored with InsightFace (buffalo_l) against five held-out reference images that are NOT in the training set — a mean cosine of 1.0 would be the same face.
18-prompt eval on Krea 2 Turbo, 768x1024, 8 steps, CFG 1, strength 1.0:
• close-up 0.88
• head & shoulders 0.89
• waist-up 0.81
• selfie 0.86
• wet hair 0.86
• full body 0.72
• profile 0.68
• mean over the 16 trigger prompts: 0.796, with 16 of 16 above 0.60
Strength curve, same prompts: 0.6 → 0.667 · 0.8 → 0.743 · 1.0 → 0.790 · 1.2 → 0.805
80 new scenes, 1024x1344, one seed each, no cherry-picking: mean 0.779, 57 of 80 at or above 0.75.
TRAINING
93 images, single source. Krea 2 Raw, rank 32 / alpha 32, 2,000 steps. Identity 0.799 on the 18-prompt held-out eval (16/16 above 0.60). Run on Krea 2 Turbo: strength 1.0–1.2, 8 steps, CFG 1, 1024x1344.
KNOWN LIMITATIONS — READ THESE
HOW TO PROMPT HER
MEASURED ON OTHER KREA 2 BUILDS (28 Sep 2026)
One constant frame — full body, symmetrical stance, white cyclorama, seed 3927885388, 1024×1536, Euler, 9 steps, CFG 1 (8 steps on Turbo), strength 1.0, no body words in the prompt. Face = similarity to her five reference pictures in this one frame (lower than the 18-prompt mean by nature). Build = hip width divided by torso length; her reference build is 0.90–1.01.
- Krea 2 Turbo: face 0.79, build 0.88 — slimmer than her reference; 0.94 with body words.
- Sick Ollie (fp8): face 0.75, build 0.86 — slimmer; body words do not change it (0.87).
- CyberRealistic Krea 2 v3.0 (fp8): face 0.77, build 0.90; 0.92 with words.
- Backdrop in this frame: a faint water texture on Turbo and printed marks on the CyberRealistic floor (the known v2.0 leak). Eyes dark brown on all three.
- v1.0 in the same frame: face 0.80 / 0.79 / 0.80, build 1.00 / 0.95 / 0.92 — fuller than her reference on Turbo.
A test build that carries her figure without any keyword measured face 0.79 / 0.77 / 0.77 and build 0.94 / 0.90 / 0.90 in the same frame, with a clean backdrop on all three. The next version is built that way.
• She leaks without the trigger. A plain prompt like "a 33-year-old South Indian woman, …" still renders her (0.80–0.84 against the references). If you want a different woman in the same scene, don't load this LoRA, or drop it to 0.6.
• The trigger word can print as text. Krea 2 renders unusual tokens onto signs, newspapers and blackboards. Avoid prompts with readable print, or expect zxrevathi written on them.
• Wide shots (v1): v1 rendered her fuller than her reference in full-body and wet-saree scenes; v2 holds her build. On v1 say "her natural figure".
• True side profiles are the weak slot (0.68). She often turns it into a three-quarter or over-the-shoulder view.
• Her face is designed to read 33. Prompting her much younger fights the model.
LICENCE AND CREDIT
This is a Derivative of Krea 2. NOTICE.txt and the Krea 2 Community License Agreement ship inside the download and travel with any copy — including its Acceptable Use Policy, content-filtering duty and revenue threshold. Not an official Krea product; not endorsed by Krea.
Adults only. She is written as a 33-year-old woman. Do not use her to depict, imply or impersonate a real person.
Made for dearai.live, where she lives as a character. Credit appreciated if you post her work.
Description
What changed: on Krea 2 Turbo, with body words in the prompt, v2.0 holds her build closer to her reference in standing, seated, side and from-behind scenes where v1.0 drifted heavier (6 of 12 scenes on a same-seed body sheet, judged by eye). Without body words she renders slimmer than her reference (build 0.88 against her 0.90–1.01 in a constant full-body frame). 125 pictures (one source, fictional). Recipe: Krea 2 Raw, rank 32/32, text encoder untrained; 2,500-step run, checkpoint step 2250; served on Krea 2 Turbo.
Measured on the same 18-prompt evaluation (16 prompts with the trigger, 2 without), against reference pictures verified by hash to be outside the training data: mean over the 16 trigger prompts 0.793, 16/16 above 0.60 (v1.0 on the same references: 0.796) — identity unchanged. Re-measured 28 Sep on five corrected references: v2.0 0.797, v1.0 0.797.
- per framing: close-up 0.86, waist-up 0.80, head & shoulders 0.86, full body 0.71, damp hair 0.86, newspaper 0.72, profile 0.70, selfie 0.85
- without the trigger she still leaks in: 0.86 / 0.82
- 30 new scenes at 1152x1440, one seed each: mean 0.826, 28 of 30 at or above 0.75
What you will see: on Turbo, body words in the prompt render as written and hold her build in full-length, seated and side scenes; without them she can come out slimmer. Superseded by v3.0, which needs no body words for her build: in her range on Turbo, a touch slimmer on the two finetunes tested (Sick Ollie, CyberRealistic). Use "photograph" in prompts, not "photorealistic" (on Krea 2 that word pulls texture toward painting).
Settings: strength 1.0 (try 0.8–0.9 for makeup looks), 8 steps, CFG 1, Euler/Simple, 1024x1344 or 1152x1440. Trigger zxrevathi + "33 years old". File sha256 1ac1aefa3afb7534…
FAQ
Comments (6)
Hi & Thanks for sharing the model!
Just started to run some photoshooting setups. What i don't get, both your LORA women lose much of their body weight when i render them in the Sick Ollie INT8 checkpoint. So they still look voluptuous, but nowhere near as thick as in your images. I'd have to add the Loraholic Skinny/Fat Slider LORA with at least +3 strength to get a comparable result.
Your prompts don't mention the body, though.
Update: They both also show completely different bodies, depending on what framing and pose is used. Like one shot is a healthy mommy, other shot is more of a bikini babe with a too-thin waist.
Additionally, there's some artifacts baked into the LORA data. For Revathi, it is water (there's often water dripping, also in some of your pics) and this can be seen very well if you make some clean white cyclorama backgrounds.
Both also don't like taking makeup (e.g. the lips don't accept lipstick much). The nose rings seem to be baked in too, though you can sometimes override them with a different one.
Thanks — this is exactly the kind of feedback I wanted, and you're right on every point.
1. Body: on Krea 2 Turbo the trigger carries her build; on Sick Ollie and other finetunes add "curvy figure, full bust, full hips, soft midriff, soft waist" and she comes back without a slider. Going on the card today.
2. Bodies changing with framing: known, and the next version tightens it — I measure the face, and from now on the body too.
3. Water: you found a real bias in v2.0's data; too many wet scenes. "dry skin, no water" helps for now; fixed in the next version. Your cyclorama test is now part of my checks.
4. Makeup: she's built with little makeup, so lipstick words are weak at 1.0; try 0.8–0.9 for makeup looks.
5. Nose stud: on purpose, it's her signature. "no nose stud" removes it about half the time.
Which Sick Ollie build and settings are you running? I'd like to test the next version on it before release.
@dearai_studio Just the regular Sick Ollie, Krea 2, INT8 checkpoint with the regular FP8 text encoder (not one of the dedicated decensored ones, for general purpose). I think there's only one VAE. Software is Forge NEO.
Thing about the added keywords like "curvy figure" to get the character look right: they are completely subjective and relative / ambiguous, so the engine renders the characters differently in each scene.
It's very very apparent here; even with / without the fat sliders, the apparent body mass changes by like 20 kilograms and more between shots. This is also why i generally don't download LORA characters which require extra keywords (not beyond e.g. "dark skin" or something about the hair) - they are never consistent.
Krea2 is still super-sensitive to these extra descriptions. Prompt "blue eyes" on any character, and the eyes become bigger in relation to the face.
@LurkingAI Thanks, that's useful. Forge Neo with the INT8 checkpoint and the standard FP8 text encoder is what I'll test on before the next release.
On keywords: agreed. Her build belongs in the trigger, not in the prompt. On Turbo that's where it is; on finetunes v2.0 leans on the words, and that's a defect, so I've pulled the "add curvy figure" advice from the card. The next version is being built to carry her build on every base without descriptors, and its card will show a consistency figure: her proportions across 12 scenes on the same seed, on Turbo and on Sick Ollie, not just a face score.
The "blue eyes" point goes on the card as you wrote it: describe the scene, not her.
@dearai_studio blue eyes was just a random example - wouldn't want your two models here with blue eyes ;)
Some models actually shift between blue, green, hazel, then it's hard to fix it without impacting the face.
@LurkingAI Ha, understood, no blue eyes on the card. Hers are dark brown and that stays with the trigger; eye colour is now one of the things I check per version before release. Thanks again.