Babes SDXL (Babes by Stable Yogi) is a Pony-based NSFW checkpoint that delivers glamour-leaning realism. For best results in 2026, run v6.5 FP16 at CFG 5 to 7, 25 to 30 steps, Euler a or DPM++ 2M Karras, native SDXL resolutions, and pair it with the creator’s PDXL positive and negative embeddings. It excels at clean, attractive figures.
We put Babes by Stable Yogi through a full week of testing across portrait, full body, and explicit NSFW scenes, on both an RTX 4090 and an RTX 3060 12GB. This review covers what the model actually does well, where it stumbles, and the exact settings that produced our best output. If you want to sample the glamour-realism look before downloading a multi-gigabyte file, try a hosted model through our browser generator first.
What Babes SDXL actually is
Despite the SDXL label many people attach to it, Babes by Stable Yogi is built on the Pony Diffusion V6 base, which is an SDXL derivative. You can find it on its Civitai model page. That Pony lineage is the single most important fact about this checkpoint, because it dictates how you prompt it. People who treat it like a vanilla SDXL model and write flowing sentences get weak, washed out results. People who use the Pony score tag system and the creator’s companion embeddings get the clean, attractive figures the model is famous for.
The current release as of 2026 is v6.5 FP16. There is also a faster DMD2 variant that runs on the LCM sampler at four steps and CFG 1.2 for users who want speed, though the standard v6.5 is what we recommend for quality. The model uses a trigger word, 99bsy99, and the creator strongly suggests adding two companion textual inversions to your prompts: Stable_Yogis_PDXL_Positives at the front of the positive prompt and Stable_Yogis_PDXL_Negatives-neg at the front of the negative.
The Babes aesthetic
Babes is not a neutral photoreal model. It has a clear bias toward conventionally attractive, glamour-styled subjects with smooth, well-lit skin. That is a feature, not a bug, if that is the look you want. It is less suited to gritty, candid, or documentary realism, where a model like RealVisXL or CyberRealistic would serve you better. Knowing this up front saves you fighting the checkpoint toward an aesthetic it was never tuned for.
Settings we verified
Here are the settings that gave us the most consistent output in testing. The companion embeddings do a lot of heavy lifting, so install them before you judge the model.
| Setting | Recommended value | Notes |
|---|---|---|
| Base model | Pony (SDXL derivative) | Use Pony LoRAs only |
| Version | v6.5 FP16 | DMD2 variant for speed at four steps |
| Sampler | Euler a or DPM++ 2M | DPM++ 2M for crisper detail |
| Scheduler | Karras | Pairs with both samplers |
| Steps | 25 to 30 | 25 is usually plenty |
| CFG scale | 5 to 7 | Start at 6 |
| Clip skip | 2 | Standard Pony setting |
| Resolution | 1024×1024, 832×1216, 1216×832 | Native SDXL aspects only |
| Trigger word | 99bsy99 | Helps lock the model aesthetic |
| Embeddings | PDXL Positives and Negatives | Strongly recommended by the creator |
The DMD2 fast variant is genuinely useful for ideation. At four steps and CFG 1.2 on the LCM sampler, you can fire off dozens of composition tests in the time one standard generation takes, then re-run your chosen seed on the full v6.5 model for the final image.
Prompting Babes correctly
Because Babes is Pony-based, it uses Danbooru style tags and the score ladder rather than natural language. Lead with the companion positive embedding, then the score tags, then your subject.
Positive:
Stable_Yogis_PDXL_Positives, score_9, score_8_up, score_7_up,
99bsy99, photo of a woman, glamour lighting, detailed skin,
soft studio light, (sharp focus:1.1), looking at viewer
Negative:
Stable_Yogis_PDXL_Negatives-neg, score_4, score_5, score_6,
(worst quality:1.2), cartoon, anime, 3d, deformed hands,
extra fingers, bad anatomy, watermark, textThings that consistently improved our results:
- Install the companion embeddings. Without them the model is noticeably flatter and the negatives are weaker.
- Keep token weights at or below 1.4. The model is already opinionated and reacts strongly.
- Use lighting tags such as
glamour lighting,softbox, orrim lightto steer the polished look it does best. - The
99bsy99trigger helps keep output on-model when your prompt drifts toward generic descriptions.
If you are new to the Pony tag grammar, the fastest way to learn is to iterate quickly. Our hosted generator lets you test prompt phrasing without a local install, and the wording you learn there ports straight into Babes locally.
Installing and running it
Babes runs in any SDXL-capable front end. Grab the safetensors file from Civitai and the two companion embeddings, then place them in the right folders.
- Automatic1111: checkpoint goes in
models/Stable-diffusion, embeddings go in theembeddingsfolder, then refresh. - Forge: same paths as A1111, and Forge handles SDXL memory more gracefully on smaller cards.
- ComfyUI: checkpoint goes in
models/checkpoints, embeddings inmodels/embeddings, loaded via the Load Checkpoint and embedding syntax.
After copying files, refresh the model and embedding lists in the UI rather than restarting. If the embeddings do not appear, confirm you placed them in the dedicated embeddings folder and not alongside the checkpoint.
Performance numbers
This is an SDXL class model, so it is heavier than SD1.5. Our measured timings:
- RTX 4090, 24GB: about 4 to 6 seconds for a 1024×1024 image at 25 steps. The DMD2 variant drops that to roughly one second.
- RTX 3060, 12GB: about 16 to 22 seconds per standard image. Very comfortable with headroom for hires fix and ADetailer.
- 8GB cards: usable in Forge with
--medvramand tiled VAE, expect 35 plus seconds. The DMD2 variant is a smart choice here to cut step count.
Where Babes shines and where it does not
After a week of testing, our honest take:
- Shines at: glamour portraits, attractive figures, soft studio and natural light, clean skin, and stylized-but-believable NSFW scenes.
- Solid at: standard NSFW anatomy, thanks to the strong Pony base underneath.
- Weaker at: gritty candid realism, older or non-idealized subjects, and complex multi-person scenes, where it can homogenize faces toward its preferred look.
- Needs help with: hands at small scale, like most SDXL models. An ADetailer hand pass or hires fix cleans this up.
The model’s strong aesthetic bias is the headline. If you want everyone to look like a polished magazine subject, Babes is excellent. If you want variety and grit, you will spend a lot of prompt effort pushing against its defaults, and you would be happier on a more neutral checkpoint. We cover those neutral alternatives in our broader checkpoint roundup, which compares Babes against the CyberRealistic and RealVisXL families directly.
We also want to be specific about prompt obedience, because it is a real strength here. When you describe a pose, an outfit, or a lighting setup in clean tags, Babes follows it more reliably than many SDXL merges, which tend to ignore secondary instructions once the main subject is set. The Pony base contributes strong concept separation, so a prompt asking for a specific pose plus a specific background plus a specific lighting style usually gets all three rather than collapsing into a generic portrait. That obedience is what makes the fixed-template batch workflow described below work so well.
Quality finishing pipeline
The finishing steps that lifted our Babes output the most:
- ADetailer on faces and hands: a single face pass at denoise 0.3 to 0.4, plus a hand pass when hands are prominent. This fixes the two most common failure points in one shot.
- Hires fix at 1.5x: with 4x-UltraSharp at denoise 0.4. Babes already produces clean skin, so the upscale mostly adds resolution rather than fixing texture.
- Light LoRA stacking: keep Pony LoRAs at 0.5 to 0.8 weight. Heavier weights fight the Babes aesthetic and muddy the skin.
- Seed locking during tuning: lock the seed while you tune CFG and steps, then unlock to explore variations once the recipe is set.
Common problems and fixes
- Flat, low contrast output: you forgot the companion embeddings or the score ladder. Add both to the front of your prompts.
- Everyone looks the same: that is the model’s aesthetic bias. Add distinguishing tags for face shape, age, and features, or switch to a more neutral checkpoint for variety.
- Broken hands: run an ADetailer hand pass or use hires fix. This is an SDXL-wide issue, not specific to Babes.
- Black images: force
--no-half-vaein A1111 or Forge. This is a VAE precision problem on certain GPUs, not a corrupt file. - Anime or cartoon leakage: confirm
anime,cartoon, and3dare in your negative, and that the negative embedding is loaded.
Comparing the standard and DMD2 variants in practice
One decision new Babes users face is which version to download. We ran both side by side and here is the practical breakdown. The standard v6.5 FP16 is the quality option. It needs 25 to 30 steps and produces the richest skin texture and lighting, and it is what you want for final images you intend to keep or publish. The tradeoff is generation time, which on an 8GB card can stretch past half a minute per image once hires fix is involved.
The DMD2 variant is the speed option. By running on the LCM sampler at just four steps and CFG 1.2, it generates almost instantly even on modest hardware. The catch is a slight reduction in micro-detail and a narrower tolerance for prompt complexity. Our recommended workflow uses both: ideate on DMD2 to find a composition and seed you like, then re-run that exact seed on standard v6.5 for the polished final. This two model loop is the single biggest productivity win we found with Babes, because it removes the slow feedback loop that makes SDXL ideation tedious on smaller GPUs.
If you only have room for one file, choose standard v6.5. The speed of DMD2 is a luxury, but the quality of the standard model is the reason to use Babes at all. A 12GB card running v6.5 hits a comfortable middle ground where ideation is fast enough that you may not need DMD2 at all.
How Babes fits a content workflow
Because Babes homogenizes toward attractive, on-model subjects, it is unusually good for producing a consistent character or a themed series. The same trigger word and embedding stack keeps faces and styling coherent across a batch, which is exactly what you want when building a recognizable persona rather than a one-off image. We leaned on this in testing: with a fixed quality block, the trigger word, and a single editable scene line, a batch of eight images held together visually far better than the same prompt on a more neutral checkpoint.
That consistency is a double edged sword. For a single persona it is a gift. For a varied gallery of different people it is a constraint you have to actively fight with descriptive tags. Match the tool to the job: reach for Babes when coherence matters, and reach for a neutral realism model when variety matters. Our checkpoint roundup lays out which models fill each role so you can keep a small, purposeful model folder instead of hoarding fifty merges that all do the same thing.
A final practical note: because Babes is Pony-based, it plays nicely with the wider Pony LoRA ecosystem. Pose LoRAs, outfit LoRAs, and style LoRAs trained for Pony all apply, as long as you keep their weights modest. That ecosystem depth is part of why Pony-derived models like Babes have stayed relevant while standalone SDXL merges have faded.
We also ran a small consistency test to see how well Babes holds a character across a series, since that is a common use case. With a fixed seed range, the trigger word, and the companion embeddings, the same described subject came back recognizably across eight generations, with consistent face structure and styling. That is meaningfully better than the average SDXL merge, where faces tend to drift between seeds. If you are building a recurring persona for a gallery or a story, this stability is one of the strongest practical reasons to choose Babes over a more neutral checkpoint, and it pairs well with a single face LoRA at low weight for even tighter consistency. Just remember to keep the prompt template fixed so the only variable between images is the scene itself.
Our verdict
Babes by Stable Yogi v6.5 is a strong, specialized NSFW checkpoint that does one thing extremely well: clean, glamour-leaning realism of attractive figures. With the companion embeddings installed and the Pony tag system respected, it produces polished output with very little fuss, and the DMD2 fast variant makes ideation painless. It is not the model for gritty or highly varied realism, but for its intended niche it is one of the best in the Pony family. Test the look first through our hosted generator, then run v6.5 locally on a 12GB or larger card for the best experience. If polished beauty is your target, Babes earns a place in your model folder.
Frequently asked questions
Is Babes SDXL really an SDXL model or is it Pony?
Babes by Stable Yogi is built on the Pony Diffusion V6 base, which is itself an SDXL derivative. So it is technically SDXL class in architecture and resolution, but it behaves like a Pony model for prompting. That means you use the score tag system and Danbooru style tags, not the natural language prompting that vanilla SDXL checkpoints prefer. Use Pony LoRAs, not base SDXL LoRAs.
Do I need the companion embeddings for Babes?
They are strongly recommended. The creator built Stable_Yogis_PDXL_Positives and Stable_Yogis_PDXL_Negatives-neg to pair with the model, and in our testing the output is noticeably flatter and the negatives weaker without them. Place them in your embeddings folder, then lead your positive and negative prompts with each one respectively. They do a lot of quality heavy lifting for very little prompt effort.
What CFG and steps work best for Babes v6.5?
Start at CFG 6 with 25 steps using Euler a or DPM++ 2M with the Karras scheduler. CFG 5 to 7 is the usable range; below 5 it gets soft and above 7 the skin can over-saturate. Twenty-five steps is usually plenty, with diminishing returns past 30. The DMD2 fast variant is different: it runs at four steps and CFG 1.2 on the LCM sampler.
What is the 99bsy99 trigger word for?
99bsy99 is the model’s trigger word that helps lock output to the intended Babes aesthetic. Adding it near the front of your positive prompt keeps results on-model, especially when your prompt drifts toward generic descriptions. It is not strictly required, but it improves consistency. Combine it with the companion positive embedding and the score ladder for the cleanest, most reliable results from this checkpoint.
Can Babes run on an 8GB GPU?
Yes, with optimization. In Forge, use the medvram flag and enable tiled VAE, and stick to native SDXL resolutions like 1024×1024. Expect roughly 35 plus seconds per standard image. The DMD2 fast variant is a smart choice on 8GB cards because four steps cuts memory pressure and time dramatically. On 12GB cards like the RTX 3060 it runs very comfortably with room for hires fix.
Why does everyone Babes generates look similar?
Babes has a strong aesthetic bias toward conventionally attractive, glamour-styled subjects. That homogenizing effect is intentional and is exactly why people choose it. To get more variety, add specific tags for face shape, age, ethnicity, and distinguishing features, and reduce reliance on the trigger word. If you need genuinely diverse, candid realism, a more neutral checkpoint like RealVisXL will serve that goal better.
What resolution should I generate at with Babes?
Use native SDXL resolutions: 1024×1024 for square, 832×1216 for portrait, and 1216×832 for landscape. Generating at SD1.5 sizes like 512×512 produces duplicated limbs and broken anatomy because the model was trained at roughly one megapixel. Once you have a good base image at native resolution, use hires fix at 1.5x to climb to higher output sizes with added detail rather than just upscaling.
How does Babes compare to CyberRealistic Pony?
Both sit on the Pony base and share the score tag workflow, but their aesthetics differ. Babes leans glamour and idealized beauty with smooth, polished skin, while CyberRealistic Pony pushes a more editorial, higher contrast realism that reads as candid. Settings are nearly identical at CFG 5 to 7 and Clip Skip 2. Choose Babes for polished beauty shots and CyberRealistic for grittier, more photographic variety.
