CivArchive
    InfluencerGram - v3.0 - Finsta
    NSFW
    Preview 149761084
    Preview 149760683
    Preview 149760694
    Preview 149760703
    Preview 149760709
    Preview 149760712
    Preview 149760716
    Preview 149760736
    Preview 149760742
    Preview 149760748
    Preview 149760751
    Preview 149760754
    Preview 149760821
    Preview 149760825
    Preview 149760833
    Preview 149760837
    Preview 149760841
    Preview 149760848
    Preview 149760859
    Preview 149760864

    InfluencerGram | MOSS-Image LoRA


    New: if you want one slim, petite body type every time and better NSFW prompt understanding, check out Slim Studio

    InfluencerGram V3 Finsta

    V3 is rebuilt from scratch: new dataset, new uncensored captions, and no more same-face.

    What's new in V3

    - Different faces again. V2 turned every woman into the same influencer. V3 faces stay close to base MOSS, with a bit more variety on top.

    - Better prompt adherence. Better than V1 and V2, on par with the model without the LoRA, and on some prompts it follows even better. Ask for her walking and she walks (V2 sat her down). Ask for baggy gym shorts and you get baggy shorts (V2 made them tight).

    - Fewer anatomy errors. Hands and arms land where you put them. In my tests base MOSS sometimes sank an arm into the body, V2 let the hand float off to the side, and V3 put it right next to her body as prompted.

    - Ethnicity actually works. Every training caption names the woman's ethnicity, so "a Black woman" or "an East Asian woman" in your prompt counts now. The set is mostly white and East Asian women, so those are the strongest.

    - Uncensored captions. The whole set was recaptioned with an uncensored captioner: face, body, nudity and anatomy all named plainly.

    - Cleaner set. Images with watermarks, TikTok and Instagram logos were cleaned out, and the set got new images with more body types, poses and outfits.

    - Not stuck on skinny anymore. V2 made everyone extremely skinny. V3 gives you curvier, more varied bodies and follows the body you prompt.

    - The phone look is back. Somewhere between V1 and V2.

    No trigger word. Load it, write your prompt, done.

    Older versions

    - V2 Slender Feed: leans hard into skinny, same face on everyone.

    - V1: the original.

    Also by me: Slim Studio, one slim body type and explicit prompts.

    V2: Slender Feed is out

    The skinny version of InfluencerGram. Slimmer bodies, no trigger word, a cleaner dataset.

    What’s better than V1:

    • No trigger word. It’s always on, so just prompt normally.

    • Slim by default. It pulls every body toward a slim frame, a narrow waist long, lean legs, camel toe, thigh gaps, sometimes even when you don’t ask for it xD.

    • Cleaner dataset. It was cut from about 1,250 to 711 hand-picked images and re-captioned with more precise body and pose descriptions.

    • Trained from scratch, not V1 trained further.

    How to use: I personally use strength 1.0, MOSS-Image SFT + the 8-step DMD2 LoRA, er_sde / simple, 8 steps, CFG 1. But you can use whatever checkpoint you want though i haven't tested it with all of them, kromoss is alright but plain SFT is my favourite. Write a full description of the body, outfit, pose, camera and setting.

    Training: 711 images, rank 32 / alpha 32, AdamW, learning rate 1e-4, 1024 px buckets, 3,000 steps.

    Want the original look on any body type? V1 is still here.

    V1

    Sharper, more authentic Instagram-style photos from MOSS-Image.

    MOSS is already really good at making Instagram photos. InfluencerGram isn't meant to replace it, just to push it further. It focuses on three things:

    • Sharpness. Clean edges, detailed faces and real per-pixel detail at full resolution. It works especially well with the 8-step DMD2 turbo LoRA, which on its own tends to look like an upscaled low-res image. With InfluencerGram the turbo output comes out crisp.

    • Authenticity. The look of real phone photos: mirror selfies, flash, casual framing, everyday rooms. Results get very close to real Instagram posts, further than the base model, KroMoss or other LoRAs I've tried.

    • Body proportions and poses MOSS doesn't know well. Petite and slim bodies, thigh gap, camel toe, and more expressive, dynamic body positions.

    It's not perfect, and I'll keep iterating on it as time goes on, especially if there's enough interest.

    How to use

    • Trigger: start your prompt with influencergram photo.

    • Prompt style: write a full natural-language description after the trigger. MOSS was trained on long captions, and so was this LoRA, so describe the body, outfit, pose, camera angle, setting and lighting.

    • Strength: 1.0. Anything from 0.8 to 1.4 makes no noticeable difference, so there's no need to fine-tune it.

    • Base model: works on MOSS-Image Base and SFT or Kromoss.

    • Turbo: stacks well with the 8-step DMD2 LoRA (both at 1.0, 8 steps, CFG 1).

    • Without turbo: CFG around 3, 30 to 50 steps. Adding watermark, username, text overlay to the negative prompt helps. (Negative prompts do nothing at CFG 1, so this tip doesn't apply to turbo.)

    Training

    • Personal Dataset: about 1,250 images, carefully curated by hand: duplicates and near-duplicates removed, and anything made with AI excluded.

    • Source: Instagram and OnlyFans. Both SFW and NSFW. All subjects are adults.

    • Captions: long, detailed natural-language captions written by Qwen3.5-9B. They describe body shape, pose, outfit, camera and setting, and quote any watermarks or text overlays, so those stay out of your images unless you prompt for them. "A watermark in the bottom right corner reads "OnlyFans.com/####"

    • Settings: trained on MOSS-Image Base with kohya sd-scripts. Rank 32 / alpha 32, AdamW, learning rate 1e-4, 1024 px buckets, 3,750 steps.

    ## Notes

    • Trained on women. Men and groups weren't a focus and are not included.

    • Watermarks and usernames appeared in a good share of the source photos. They were captioned so the LoRA treats them as separate from the look, but a faint corner text can occasionally slip through.

    Description

    Different faces. V2 turned every woman into the same influencer. V3 faces stay close to base MOSS, with a bit more variety on top.

    Better prompt adherence. Better than V1 and V2, on par with the model without the LoRA, and on some prompts it follows even better. Ask for her walking and she walks (V2 sat her down). Ask for baggy gym shorts and you get baggy shorts (V2 made them tight).

    Fewer anatomy errors. Hands and arms land where you put them. In my tests base MOSS sometimes sank an arm into the body, V2 let the hand float off to the side, and V3 put it right next to her body as prompted.

    Ethnicity actually works. Every training caption names the woman's ethnicity, so "a Black woman" or "an East Asian woman" in your prompt counts now. The set is mostly white and East Asian women, so those are the strongest.

    Uncensored captions. The whole set was recaptioned with an uncensored captioner: face, body, nudity and anatomy all named plainly.

    Cleaner set. Images with watermarks, TikTok and Instagram logos were cleaned out, and the set got new images with more body types, poses and outfits.

    Not stuck on skinny anymore. V2 made everyone extremely skinny. V3 gives you curvier, more varied bodies and follows the body you prompt.

    The phone look is back. Somewhere between V1 and V2.

    FAQ