Sweeti Girl — Pornsuit Edition

Model type: Stable Diffusion 1.5 Full Fine-tune (UNet only) Dataset: 7,175 images at 640×960 resolution Training steps: 280,000 Hardware: RTX 4080 Super (16GB VRAM)

Overview

Sweeti Girl is a full UNet fine-tune built on Stable Diffusion 1.5, trained on 7,175 carefully curated portrait images at 640×960 resolution over 287,000 steps on a single RTX 4080 Super. The CLIP text encoder was intentionally left untouched, meaning the model retains full prompt responsiveness from the SD 1.5 base while the UNet carries the entire aesthetic signature of the training data. The result is a model that behaves predictably with standard prompts while delivering a consistent and distinctive photographic quality across an exceptionally wide range of subjects, styles, and scenarios.

This is not a concept model or a character LoRA. It is a general-purpose photorealistic portrait model with NSFW fully active, designed to generate high-quality images of women across diverse ethnicities, lighting conditions, environments, and clothing styles without requiring special trigger words or complex prompt engineering.

Racial Generalization — A Core Differentiator

The most technically significant feature of Sweeti Girl is its genuine racial generalization capability. Across all tested batches, the model produced convincing and natural-looking subjects of European, Mediterranean, Asian, Latina, and mixed-heritage appearance, all with equal quality and without any drift correction in the prompts.

This behavior is a direct consequence of the UNet-only training approach. Because the CLIP encoder was not fine-tuned, the model's semantic understanding of ethnicity-related terms remained exactly as SD 1.5 left it. The UNet learned a generalized aesthetic quality — skin texture, lighting response, anatomical proportions, hair behavior — that applies equally across all ethnic groups rather than collapsing toward a single dominant type as many fine-tuned models do.

In practical terms this means that prompting asian woman, european woman, latina woman, or leaving ethnicity unspecified all produce results of equivalent quality with no visible degradation or artifacts. Each ethnic variant feels like an intentional artistic choice rather than a model failure or workaround. This level of generalization is rare in SD 1.5 fine-tunes and represents one of the strongest arguments for using this model over more specialized alternatives.

Facial Quality

Facial rendering is where Sweeti Girl most clearly distinguishes itself from the SD 1.5 baseline and from most community fine-tunes. The model consistently produces faces with individual pore-level skin texture, realistic subsurface light scattering, naturally varied microexpressions, and correct anatomical proportions across different angles and distances.

Close-up portrait tests at extreme focal lengths revealed skin texture detail that rivals SDXL fine-tunes on equivalent prompts. Individual eyelashes are rendered as separate strands rather than painted-on masses. Iris detail includes visible pupil depth, limbal ring variation, and specular catchlights positioned correctly relative to the light source in the scene. Eyebrow hair direction and density are consistent across seeds within the same prompt context.

The model handles smiling expressions unusually well. Many fine-tuned SD 1.5 models struggle with natural smiles, producing either frozen or uncanny results. Across outdoor park and restaurant terrace batches, Sweeti Girl generated warm, organic smiles with correct cheek muscle engagement and natural teeth rendering without requiring specific smile-related prompting.

Ethnic facial structure adapts correctly to context. Asian subjects show appropriate bone structure variation — higher cheekbones, softer jaw angles, monolid or double-lid eye structure — without any of the blending artifacts that occur when a model has learned only a single dominant face type. European subjects show the same structural intelligence applied to a different template.

Hair Rendering

Hair is consistently one of the most technically demanding elements in portrait generation, and it is one of Sweeti Girl's strongest areas. The model was tested across platinum blonde, natural blonde, dark brown, black, auburn, and red hair at lengths ranging from short pixie cuts through shoulder-length bobs to very long flowing styles.

In all cases the model produces individual strands rather than painted hair masses. Backlit shots — particularly the golden hour field sequences and rooftop sunset batches — show correct light transmission through hair, with the translucency varying naturally based on strand density and distance from the light source. The auburn and copper hair batches produced particularly impressive results, with warm-to-cool color gradients across individual sections of hair that respond to the ambient light color in the scene.

Black hair on Asian subjects rendered with deep, natural gloss rather than the flat matte black that many models produce. Platinum blonde hair consistently maintained the correct optical behavior of very light hair — high specular response, visible strand separation, natural volume without the plastic sheen that lower-quality models generate.

Hair in motion — wind-blown styles in outdoor and rooftop batches — showed physically plausible movement with strand cohesion rather than random floating. The model learned that hair moving in wind maintains directional consistency and that individual strands follow adjacent ones rather than moving independently.

Lighting Mastery

Lighting is arguably the single area where Sweeti Girl most exceeds expectations for an SD 1.5 model. Across all tested batches, the model demonstrated the ability to render at least eight distinct lighting scenarios with photographic accuracy.

Window light with geometric shadows. The indoor batches using window light as the primary source produced some of the technically best images in the entire test set. Shadow geometry from window frames, venetian blinds, and foliage fell across faces and backgrounds with correct angular behavior relative to implied light source position. The transition from highlight to shadow on skin followed physically correct gradients rather than the hard or blurry edges that indicate model uncertainty.

Golden hour backlight. Outdoor field and park batches demonstrated exceptional mastery of backlit scenes. The model correctly renders the halo effect of strong backlight on light-colored hair, the warm color shift that occurs when sunlight passes through skin and hair at shallow angles, and the slight lens flare characteristics of a real camera pointing toward a low sun.

Urban night bokeh. Multiple night-time city batches produced correctly rendered circular bokeh from street and building lights, with the bokeh size varying naturally based on implied distance and lens aperture. Light color variation across different light sources in a single frame — warm sodium street lights alongside cool LED shop fronts — was handled with consistency.

Natural overcast outdoor light. Park and outdoor café batches in diffuse light conditions showed correct shadowless illumination with subtle color temperature variation between sky light and reflected ground light.

Mediterranean direct sunlight. The brick wall fashion batches and noir series demonstrated the model's ability to render hard, angular shadows from direct overhead or lateral sunlight, including the slight color warming that occurs in shadows relative to directly lit areas.

Interior ambient. Lingerie and bedroom batches in soft interior light showed correct wraparound shadow behavior on three-dimensional forms, with natural falloff gradients that define volume convincingly.

NSFW Performance

The NSFW capability is fully active and performs at a level consistent with the overall photographic quality of the model. Anatomy is rendered with correct proportions and natural variation rather than the exaggerated or uniform appearance that many NSFW models converge toward. Body proportions across different poses maintain internal consistency — the model understands how the same body changes in appearance between standing, seated, and turned positions.

Lingerie and intimate apparel is rendered with particular care. Lace fabrics show correct transparency behavior with underlying skin tones visible through the pattern rather than the fabric being treated as an opaque solid with a texture overlay. Garter belts and stockings maintain structural coherence with correct attachment points and natural stocking tension. Metallic hardware — buckles, clasps, rings — catches light correctly relative to the scene's primary light source.

Skin rendering in NSFW contexts maintains the same pore-level texture quality observed in clothed portraits. There is no visible quality discontinuity between face and body, which is a common failure mode in models where the NSFW capability was added as a separate training stage rather than integrated throughout.

Clothing and Material Rendering

Across all tested clothing types, the model demonstrates strong understanding of how different materials behave visually.

Ribbed knitwear — tested extensively across multiple batches — shows correct cable structure with light catching the raised ribs and shadow falling in the recessed channels. The fabric stretches and compresses plausibly around the body without breaking into artifacts at areas of tension.

Satin and silk surfaces produce correct specular highlights with the elongated, soft-edged reflection characteristic of these materials rather than the sharp point specular of rougher surfaces. The red satin blazer in the brick wall fashion batch was particularly well-rendered, with the highlight moving correctly across the garment in a way that implied correct understanding of the fabric's drape and curvature.

Velvet in the forest batches showed the distinctive nap-dependent light behavior of this material, with areas of the fabric facing toward the light appearing lighter and those facing away appearing darker in a pattern that follows the fabric's three-dimensional structure.

Floral printed fabrics maintained pattern coherence without the geometric distortion that occurs when a model treats print as a flat texture applied to a shape rather than as a property of the fabric itself.

Environment and Background Rendering

Background and environment rendering is competent across all tested scenarios. Urban environments — city streets, rooftop city views, café terraces, night-time shopping districts — are rendered with spatial coherence and appropriate population density. People in backgrounds are correctly scaled and positioned and do not show the distortion artifacts that often indicate a model struggling with secondary figure generation.

Natural environments — forests, fields, parks, lakesides — produce convincing vegetation bokeh at middle distances with correct depth-of-field behavior. The foggy forest sequence produced one of the most atmospherically successful images in the entire test set, with correctly graded depth haze and subtle tonal variation between fog layers at different distances.

The Grand Canyon sequence exposed the model's primary background weakness: at high detail levels in very large, patterned environments, a tiling artifact appears where geological formations repeat with unnatural regularity. This is a known SD 1.5 limitation rather than a model-specific failure and does not appear in any other tested environment type.

Interior environments — hotels, apartments, lobbies, corridors — show correct perspective and architectural coherence. Window-lit interiors show the correct relationship between bright window openings and the relatively darker interior walls around them.

Character Consistency Across Seeds

One of the key practical qualities of a general portrait model is whether it maintains a recognizable aesthetic identity across different seeds, prompts, and scenarios. Sweeti Girl performs well on this measure. Images generated with widely different prompts — a platinum blonde in a field versus an Asian woman in a Tokyo street versus a Mediterranean brunette in a lingerie interior — all feel like they come from the same model rather than from different sources. The consistent quality of skin rendering, the similar approach to lighting, and the stable anatomical understanding tie diverse outputs together into a coherent body of work.

Within more constrained prompt sets — same general description, varied seeds — the model produces the expected range of natural variation while keeping core identity stable. Faces vary naturally in the same way that different photographs of a similar-looking person would vary, rather than producing either clones or unrelated individuals.

Technical Architecture Notes

The decision to train only the UNet while leaving the CLIP text encoder at SD 1.5 baseline has three observable consequences. First, prompt adherence is strong and predictable. Standard photography and portrait prompts behave exactly as experienced SD 1.5 users would expect, with no vocabulary shifts or prompt translation required. Second, there is no evidence of catastrophic forgetting in any tested concept area. Environments, objects, clothing types, and activities that were not part of the training data render correctly because the model's conceptual vocabulary was not modified. Third, the model stacks cleanly with LoRA models trained on SD 1.5, since neither the text encoder weights nor the UNet architecture were changed in ways that would create incompatibility.

The 640×960 training resolution — a 2:3 portrait aspect ratio — means the model's strongest output is in portrait-oriented crops at or near this ratio. Landscape crops and square crops are usable but represent a slight departure from the model's native orientation. Upscaling via hires fix from native resolution produces clean results consistent with the base model's upscaling behavior.

Sweeti Girl Pornsuit Edition — trained on RTX 4080 Super — SD 1.5 UNet fine-tune

⚠️ CONTENT WARNING (NSFW):

This merge is uncensored and capable of generating high-quality explicit NSFW content and nudity. It has a tendency towards revealing clothing in casual settings.

• For SFW results: Strong negative prompts are highly recommended (e.g., nude, nipples, explicit, nsfw).

⚠️ LICENSE & PERMISSIONS (READ BEFORE DOWNLOADING)

1. PERSONAL USE ONLY This model is provided free of charge for Personal, Non-Profit, and Research use only. You may use it to create images for your personal portfolio.

2. STRICTLY NO REDISTRIBUTION

• ❌ DO NOT re-upload this file to Civitai, Hugging Face, or any other platform.

• ❌ DO NOT host this model on third-party generation services (e.g., Tensor.art, Mage.space, Telegram Bots).

3. COMMERCIAL RESTRICTIONS Using this model or its outputs for commercial revenue (Influencers, Ads, Stock Photos) without a license is PROHIBITED.

💼 COMMERCIAL SERVICES & COMMISSIONS

I do not sell the model file for commercial use. Instead, I offer premium AI solutions for brands and agencies:

• ✨ Exclusive AI Influencers: I create and manage consistent digital personas for Instagram/Social Media.

• 🏢 Corporate B2B LoRAs: Custom training for brand identity and mascots.

• 📸 High-End Image Packs: Monthly content packages for your brand.

To hire me for professional AI Modeling services: 📩 Contact: [[email protected]]

I recommend using the Adetailer extension.

Use this extension to fix hand errors:

https://github.com/licyk/advanced_euler_sampler_extension

Use these recommended settings for generation:

Sampling method: Euler_Max

Sampling steps: 30-50

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: DPM++ 2M

Sampling steps: 18-30

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: Restart

Sampling steps: 30-50

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: Kohaku_LoNyu_Yog

Sampling steps: 30-50

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: Euler_Smea_Dy

Sampling steps: 18-50

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: Euler a

Sampling steps: 18-50

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: LCM

Sampling steps: 18-30

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: DDPM Karras

Sampling steps: 18-30

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: DPM++ SDE Karras

Sampling steps: 18-30

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

Sampling method: DPM++ 2M SDE Karras

Sampling steps: 18-30

CFG Scale: 2.0 - 7.0

Skip clip: 1-2

(CyberRealistic_Negative-neg), deformed, bad anatomy, bad hands, missing fingers,

extra fingers, mutated hands, poorly drawn hands,

blurry face, out of focus face, cartoon, anime,

illustration, painting, drawing, 3d render,

watermark, text, signature, oversaturated, bad neck,

plastic skin, doll, unrealistic, low quality,

flat lighting, overexposed, nude, asian, chinese, japansese