Best SD 1.5 Uncensored Models on Civitai (2026 Rankings)
SD 1.5 still dominates local uncensored generation for one reason: nothing else has its LoRA library. Thousands of style LoRAs, character LoRAs, and pose packs all target SD 1.5 conditioning - and they all work with these checkpoints out of the box.
You don't need a 12 GB GPU. You don't need SDXL. You need a proven checkpoint, correct settings, and a local frontend that doesn't filter your prompts. This page ranks the six SD 1.5 checkpoints that consistently deliver for NSFW work on Civitai, with real version numbers, tested CFG values, and honest opinions on where each one excels or falls short.
Every model here runs on 4–6 GB VRAM in Forge or ComfyUI. I've tested each one at 512×768 native resolution with hires fix for final output. The rankings reflect quality, consistency, and how well each model responds to the massive SD 1.5 LoRA ecosystem.
The Models
1. Realistic Vision V6.0 B1
Top PickThe default SD 1.5 photoreal checkpoint. 2.2M downloads, exceptional skin texture, and full LoRA compatibility make it the obvious starting point.
Architecture: SD 1.5 · VRAM: 4 GB+ · Best for: Photorealistic portraits and full-body
Open on Civitai →Cleaner skin than Realistic Vision at low CFG. Opinionated - wants simple prompts and refuses to cooperate with keyword stacking.
Architecture: SD 1.5 · VRAM: 4 GB+ · Best for: Clean studio photorealism
Open on Civitai →The 'less is more' model. Skip quality keywords entirely - describe your scene naturally and let the model handle lighting and detail.
Architecture: SD 1.5 · VRAM: 4 GB+ · Best for: Natural candid photorealism
Open on Civitai →13.7M downloads for a reason - handles illustration, semi-real, and fantasy from one checkpoint. Not the photoreal king, but the flexibility king.
Architecture: SD 1.5 · VRAM: 4 GB+ · Best for: Multi-style versatility
Open on Civitai →Maximum detail output that upscales beautifully. Needs more pipeline setup (embeddings + ADetailer) to reach peak quality.
Architecture: SD 1.5 · VRAM: 4 GB+ · Best for: High-detail with ADetailer workflows
Open on Civitai →Breaks the 512px convention - handles 832×1216 natively. Fewer hires fix passes needed. Stick with v5 specifically for SD 1.5 LoRA compat.
Architecture: SD 1.5 · VRAM: 4 GB+ · Best for: Native high-res SD 1.5 generation
Open on Civitai →The Quick Answer
Key Takeaway - May 2026
Realistic Vision V6.0 B1 is the best overall SD 1.5 uncensored checkpoint for photorealistic NSFW - 2.2 million downloads, excellent skin texture, and it responds perfectly to the entire SD 1.5 LoRA library. If you want natural-looking results with minimal prompting, start here with CFG 5, DPM++ SDE Karras, 25 steps, 512×768.
Runner-up: CyberRealistic Final for cleaner studio-lit skin. Versatility pick: DreamShaper 8 if you need both illustration and photo styles from one checkpoint.
Ranked Model Breakdown
1. Realistic Vision V6.0 B1
| Architecture | VRAM | Best For |
|---|---|---|
| SD 1.5 | 4 GB+ | Photorealistic portraits, full-body, natural lighting |
The photoreal king of SD 1.5. SG_161222's Realistic Vision line has 2.2 million downloads and 52k likes for good reason - it produces convincing skin texture, natural lighting falloff, and anatomically coherent bodies without heavy prompting.
Version to grab: V6.0 B1 (VAE baked in). Skip older versions unless you have workflows locked to V5.1.
Settings that work:
- Sampler: DPM++ SDE Karras
- Steps: 25–30
- CFG: 3.5–7 (start at 5)
- CLIP Skip: 1
- Resolution: 512×768, then hires fix to 2x
- Negative: keep it short - "bad anatomy, bad hands, blurry" is enough
The V5.1 Hyper variant runs at 4–6 steps with CFG 1.5–2.0 if you want speed over fine control. Great for batch exploration, less precise for final renders.
2. CyberRealistic Final
| Architecture | VRAM | Best For |
|---|---|---|
| SD 1.5 | 4 GB+ | Clean studio photorealism, even skin tones |
CyberRealistic punches above its download count. The "Final" version (model 15003) delivers strikingly clean skin with fewer artifacts than Realistic Vision at comparable settings. VAE is baked in - one less thing to configure.
Settings that work:
- Sampler: DPM++ SDE Karras
- Steps: 30
- CFG: 5 (don't go higher - realism degrades fast)
- Resolution: 512×768
- Upscale: 2x with 4x_NickelbackFS_72000_G, denoising 0.3
- Negative: "lowres, bad anatomy, bad hands, text, error, blurry"
The tradeoff: CyberRealistic is opinionated. It wants simple prompts and low CFG. Fight it with heavy keyword stacking and you'll get waxy, overprocessed results.
3. epiCRealism Natural Sin RC1
| Architecture | VRAM | Best For |
|---|---|---|
| SD 1.5 | 4 GB+ | Natural candid photorealism, minimal prompting |
epiCRealism (model 25694) is the "less is more" checkpoint. Skip keywords like "masterpiece, best quality" - they actually hurt output here. This model excels at natural-looking imagery where lighting, shadows, and skin detail emerge without being prompted explicitly.
Settings that work:
- Sampler: DPM++ SDE Karras or DPM++ 2M Karras
- Steps: 25+
- CFG: 5 (higher values kill the realism)
- Resolution: 512×768
- Upscale: 2x with 4x_NMKD-Superscale-SP, denoising 0.35
Critical tip: Avoid "1girl" style anime tokens. They push epiCRealism toward uncanny territory. Describe the scene naturally - "woman standing in doorway, soft afternoon light" beats "1girl, solo, standing."
4. DreamShaper 8
| Architecture | VRAM | Best For |
|---|---|---|
| SD 1.5 | 4 GB+ | Versatile all-rounder - illustration, semi-real, fantasy |
DreamShaper 8 (model 4384) has a staggering 13.7 million downloads. It's not the most photorealistic option - that's not its job. DreamShaper handles illustration, semi-realistic, and fantasy styles from a single checkpoint. If you run diverse LoRAs across multiple styles, this is your base.
Settings that work:
- Sampler: DPM++ 2M SDE Karras
- Steps: 20–30
- CFG: 7–9 (tolerates higher values than photoreal models)
- CLIP Skip: 2
- Resolution: 512×768 or 512×832
- Negative embedding: BadDream (download from Civitai)
Why it ranks 4th: For pure photorealistic NSFW, the top three beat it. But if you switch between anime, illustration, and semi-real styles daily, DreamShaper's flexibility is unmatched at SD 1.5.
5. AbsoluteReality v1.8.1
| Architecture | VRAM | Best For |
|---|---|---|
| SD 1.5 | 4 GB+ | High-detail photorealism with ADetailer workflows |
AbsoluteReality (model 81458) is the detail maximizer. It responds exceptionally well to ADetailer for face enhancement (keep denoising under 0.3) and produces sharp, high-detail output that upscales cleanly.
Settings that work:
- Sampler: DPM++ SDE Karras
- Steps: 25–30
- CFG: 4.5–7
- CLIP Skip: 2
- Negative embeddings: BadDream + UnrealisticDream (weight 1.2–1.5)
- Hires fix: 8x_NMKD-Superscale_150000_G, denoising 0.45, 20 steps
Trade-off: Needs more post-processing setup (embeddings, ADetailer) to reach its peak. Out-of-the-box results are good; with the full pipeline, they're excellent.
6. PerfectDeliberate v5
| Architecture | VRAM | Best For |
|---|---|---|
| SD 1.5 | 4 GB+ | Higher-res native generation, versatile merge |
PerfectDeliberate v5 (model 24350) breaks one SD 1.5 convention: it handles resolutions above 512 natively. Settings suggest 832×1216 up to 1024×1536 - unusual for SD 1.5 and useful if you want larger compositions without relying entirely on hires fix.
Settings that work:
- Sampler: Euler a or DPM++ 2M Karras
- Steps: 20–50
- CFG: 5–8
- CLIP Skip: 2
- Resolution: 832×1216 (native, no hires fix needed for many compositions)
Why it's interesting: 246k downloads with 4,373 overwhelmingly positive reviews. The higher-res native capability means fewer two-pass workflows. The downside: newer versions (v8+) moved to Illustrious architecture, so stick with v5 specifically for SD 1.5 LoRA compatibility.
Quick Comparison
| Rank | Model | Downloads | CFG | Steps | Strength |
|---|---|---|---|---|---|
| 1 | Realistic Vision V6 | 2.2M | 3.5–7 | 25–30 | Best photoreal skin/anatomy |
| 2 | CyberRealistic Final | - | 5 | 30 | Cleanest studio lighting |
| 3 | epiCRealism Natural Sin | - | 5 | 25+ | Most natural candid look |
| 4 | DreamShaper 8 | 13.7M | 7–9 | 20–30 | Style versatility king |
| 5 | AbsoluteReality v1.8.1 | - | 4.5–7 | 25–30 | Best detail + ADetailer combo |
| 6 | PerfectDeliberate v5 | 246K | 5–8 | 20–50 | Native high-res SD 1.5 |
Who Should Use What
- "I want photorealistic NSFW with minimal setup" → Realistic Vision V6. Install Forge, drop in the checkpoint, generate.
- "I want the cleanest skin texture possible" → CyberRealistic Final. Low CFG, simple prompts, let the model do the work.
- "I shoot for natural/candid aesthetics" → epiCRealism. Describe scenes, not quality keywords.
- "I use different styles daily" → DreamShaper 8. One checkpoint, multiple personalities.
- "I want maximum detail and don't mind ADetailer setup" → AbsoluteReality v1.8.1.
- "I hate hires fix and want native higher-res" → PerfectDeliberate v5 at 832×1216.
Frontend recommendation: Forge for WebUI-style workflow, ComfyUI for node automation. LocalForge AI bundles both approaches without Python dependency management if you want generation running in under five minutes.
Bottom Line
SD 1.5 uncensored checkpoints aren't going anywhere. The LoRA ecosystem is too deep, the VRAM requirements are too friendly, and the community knowledge base is too mature. Realistic Vision V6 is the default recommendation - but CyberRealistic and epiCRealism both beat it in specific scenarios.
Pick one checkpoint. Learn its CFG sweet spot. Lock a seed. Change one variable at a time. That workflow produces better results than downloading fifteen models and never learning any of them properly.