Feel free to join in our community discord: https://discord.gg/YvSuxtdgH
RawGirlSpicy -- v1
The spicy sibling of RawGirl — same raw-photo DNA, now fully NSFW.
Real-photo energy, zero AI gloss. Deep focus, natural skin, real lighting, that grainy "already exists on someone's camera roll" look — carried straight over into explicit territory. No editorial polish, no plastic rendering, no fake bokeh.
Full NSFW. Trained end-to-end for adult content, so it doesn't need a separate NSFW LoRA propping it up.
✨ What it does
Handles every prompt style. Verbose, tag-soup, natural language, short and blunt — it stays coherent and doesn't fall apart.
Realism is baked in. Even used completely on its own, RawGirlSpicy is trained on the same realism foundation as RawGirl. It's not just an NSFW patch — it stands alone.
Stacks with RawGirl for the best results. Running both together pushes realism noticeably further: better texture, better skin, better light. If you want the most photographic output, run the pair.
Plays nice with character LoRAs. Layer as needed.
⚙️ How to use
Recommended strength: 1.2 — this is the value used in the reference workflow and the sweet spot for the model.
Works great on its own at 1.2. Dial up or down to taste depending on your checkpoint.
Combined with RawGirl: run both at 1.2 as per the workflow. Adjust the balance between the two if you want to control how hard the realism hits versus how far the spicy side leans.
For best adherence, follow the prompt structure used in the sample images posted with this version.
💡 Tips
Lean into the realism: describe clutter, real lighting, specific fabrics, imperfections. The more your prompt reads like a photo that already exists, the better it hits.
Deep depth of field / sharp background prompts work best — this is a no-bokeh model, same as RawGirl.
v2 — What's new
Better textures. Skin, fabric, and surface detail hold up more convincingly — sharper and more photographic at full resolution.
More amateur feel. Pushed further into that genuine "someone actually shot this" look — less polished, more real camera-roll energy.
Improved concepts. Broader and more reliable understanding across scenarios and setups.
Prompt-friendly. Seems to work well with every prompt style — verbose, tag-soup, natural language, or short and blunt.
⚙️ Usage changes
Recommended strength is now 1 (down from 1.2). That's the new sweet spot for v2 — adjust to taste depending on your checkpoint.
Still stacks cleanly with RawGirl and character LoRAs — layer as needed.
For best adherence, follow the prompt structure used in the sample images posted with this version.
Description
v3 — What's new
10x larger. The LoRA has been scaled up tenfold in size compared to previous versions — far more capacity to capture detail, variance, and the nuances of realism.
Maximum realism achieved. This version pushes realism to its ceiling — texture, lighting, imperfections, all pushed as far as the LoRA can render.
Concepts expanded to the max. Broader and more consistent coverage of scenarios, poses, and situations than any previous version.
Likely the final chapter of the saga. This build represents the endpoint of the RawGirl project — there's probably not much further to push on this path.
⚙️ How to use
Recommended strength: 0.9.
Compatible with character LoRAs — stacks cleanly to give your character that real-photo look.
Also compatible with other LoRAs (including RawGirlSpicy) — balance to taste depending on the result you're going for.
📸 A note on the sample images
All sample images are raw output — no upscaler, no post-processing, no editing — so everything you see is a direct result of the LoRA itself.
Generated with a basic workflow, no special tricks: what you get with a standard setup is exactly this.
FAQ
Comments (13)
how come v3 is several gigs?
10x larger. The LoRA has been scaled up tenfold in size compared to previous versions — far more capacity to capture detail, variance, and the nuances of realism.
even 1GB would be overkill unless it knows several different concepts. More is just too much.
What does the stripped version look like? Manageable?
@lost_moon Normally I would concur with this, but in this case, at least on the technical side, it really is an extremely dense lora.
Resizing to the standard rank 32 (218mb) caps in all layers and has 64.6% retention, which is extremely low (I'd expect at least 80% for x10 reduction). It also took an extremely long time for some reason. I am attempting rank 128, but it would appear to be doing something with those extra weights. Whether that is overfitting / noise / content / minutiae, I cannot say. Can send the compression if you'd like to test.
Rank 128, 900mb, 80.73%. Dense all the way. Just a handful of layers below 128.
People would gladly use a 30GB model but as soon as a Lora is larger than a crappy Rank 32 200mb file they get upset
@Couldvebeenworse no, the majority would not want to use a 30GB model. Hardware constraints are a real issue. Processing speed and storage as well. We're not talking about a 200mb file here.
This is almost all female-focus in terms of the training data, right? I assume so because the name RawGirl does seem specific. Just checking to be sure, though.
lmao 3 gig? i wonder if it's really necessary, i'm downloading anyways
jajajaj que haces gordo AI
If anyone has tried it would you say its better than realism engine v3.1 for NSFW?
Can you explain how to use this? I don't get it. Do I just type “Amateur iPhone shot” at the beginning of the prompt, and that's it? Can you suggest some effective phrases for getting that Lora style?