Pro Version of Realism Pony V4-V5-V6 now. Get the Pro version of my models here
Onsite generations are permanently available on these models:
👉 Realism_By_Stable_Yogi V3: https://civarchive.com/models/166609?modelVersionId=992946
Realism by Stable Yogi Pony V6.5
V6.5 is here — and you all helped build it.
Real thank-you to everyone who pushed V6 hard, sent feedback, and posted the broken hands. V6.5's fix list literally came from you. Anatomy, hand grips, expressions, twin-tails, full-body proportions, isolated objects, painterly style separation, hair color consistency — all worked on this round.
Trigger Word
99rbsy99 — add this to every prompt for the V6.5 realism style. Place it at the END of your tag list for soft activation, or earlier for stronger effect.
Compatible with my character LoRAs (which use 99bsy99) — they stack cleanly without conflict. Use both together for a character rendered in V6.5 realism.
All Variants in This Release
Seven variants ship today, covering everything from 4 GB CPU setups to 24 GB workstations.
FP32 (safetensors, around 13 GB)
Maximum precision. Research and production work. Best for 24 GB+ cards.
FP16 (safetensors, around 6.5 GB)
The default. Best quality and speed balance for most users.
BF16 (safetensors, around 6.5 GB)
Same size as FP16, slightly faster on RTX 3000+ with native BF16 support.
FP8 Scaled (safetensors, around 3.2 GB)
Near-FP16 quality at half the VRAM. Native in Forge and ComfyUI. Great for 8 GB cards.
DMD2 Merge (safetensors, around 6.5 GB)
FP16 with DMD2 distillation LoRA pre-merged. 4-step generation. LCM sampler, CFG 1.2. Fastest path for any card.
Q8_0 GGUF (around 3.9 GB)
8-bit quantized. Near-FP16 quality. For 12+ GB cards in GGUF workflows.
Q4_0 GGUF (around 2.7 GB)
4-bit quantized. Smallest file. Makes SDXL actually run on 6–8 GB entry-level cards.
Quick Pick by Your VRAM
24 GB+ (3090, 4090, 5090, A6000) — FP16 or BF16. No reason to compress.
12–16 GB (3060 12GB, 4070, 4080) — FP8 Scaled or Q8_0 GGUF. Near-FP16 quality with headroom for LoRAs.
8–12 GB (3060, 4060 Ti, 2080) — FP8 Scaled or Q8_0 GGUF. Solid quality, comfortable VRAM use.
6–8 GB (3050, 2060, 1660) — Q4_0 GGUF. Smallest file, makes SDXL actually work on entry-level cards.
CPU only or 4 GB cards — Q4_0 GGUF in ComfyUI-GGUF. Slow but functional.
DMD2_Fp16 variant. 4 steps instead of 25–30.
Recommended Settings
For FP32, FP16, BF16, FP8 Scaled, and GGUF variants:
Sampler — DPM++ 2M Karras, Euler a, or Restart
Steps — 25 to 30
CFG — 4 to 7
Resolution — Native SDXL (1024×1024 or aspect-ratio buckets)
For DMD2 specifically:
Sampler — LCM
Steps — 4 (not 25+)
CFG — 1.2 (not 7)
Result — Comparable quality to a 25-step generation in roughly 1/6 the time
Quants Explained — Which File Do I Pick?
If you've ever seen FP16, BF16, FP8, Q4, Q8 and just downloaded the biggest one, this section is for you.
What's a quant
? Same model, smaller file. Weights are compressed so they fit on less VRAM. Some quality loss vs FP16, but smart compression (Q8_0) is so close you won't see a difference in normal use.
Quality Ladder
FP16 ≈ BF16 ≈ Q8_0 > FP8 > Q4_0. Above Q4_0 the differences are basically invisible in normal generation.
About Speed
Smaller quants are NOT always faster. Generation speed is mostly compute-bound on most cards — quants help with VRAM fit, not raw iterations per second. Where they DO help speed: avoiding system-RAM offload, which is what kills speed on small cards when the model doesn't fit.
Three Reasons to Use a Quant
VRAM fit. A 6 GB card cannot load a 6.5 GB FP16 SDXL — your UI will try to offload to system RAM and generation crawls to under 0.1 iterations per second. A Q4_0 fits with room to spare.
Speed via avoiding offload. Once a model fits in VRAM, speed depends on your card's compute, not file size. But the second it doesn't fit, speed drops 10 to 100 times. Quants are insurance against that cliff.
More room for LoRAs, ControlNet, hires fix. Even if FP16 technically fits, loading a couple of LoRAs and a ControlNet on top can push you over. Q8_0 leaves you 2–3 GB of headroom for the rest of your stack.
How to Load GGUF Files
GGUFs need a loader, since most UIs don't natively support them yet.
For ComfyUI — install the ComfyUI-GGUF custom node:
https://github.com/city96/ComfyUI-GGUF
For Forge or Forge Neo — install my Forge SDXL GGUF extension:
https://github.com/brandulateai/sd-forge-sdxl-gguf-brandulateai
After installing, GGUFs load straight from the standard checkpoint dropdown. No external module picker, no extra setup.
All my GGUFs are bundled (UNet + CLIP-L + CLIP-G + VAE in one file) so they load without picking separate components.
Pro Version Available
This is the standard version of V6.5. The Pro version is trained on more data for longer, producing a more polished and refined output. Get the Pro version of my models here
Found Anything Off?
Drop it in the comments or on Discord. V7's fix list starts now.
Want to contribute to checkpoint feedback, signup here Studio.Brandulate
Loving this model? Get the Pro version of my models here for exclusive perks and early access to unique resources.
To discuss custom LoRa's or models, feel free to connect on Discord.
👍 Like this model to keep me motivated and inspired to create more!
💬 Drop a comment and let me know what you'd love to see next.
🌟 Review this model to help me improve and make even better creations.
🔔 Hit that notification bell to stay updated with my latest models and updates!
Important Usage Tips
Add Stable_Yogis_PDXL_Positives at the beginning of your prompt section.
Add Stable_Yogis_PDXL_Negatives-neg at the beginning of your negative prompt section.
Description
V6.5 Speed & Low-VRAM Variants
Companion to the main FP16 / BF16 / FP8 release. Same V6.5 model — optimized for low-VRAM cards and speed-first workflows.
This Post
DMD2 Merge (safetensors, around 6.5 GB) — FP16 with DMD2 distillation LoRA pre-merged. 4-step generation. LCM sampler, CFG 1.2. Fastest path for any card.
Q8_0 GGUF (around 3.9 GB) — 8-bit quantized. Near-FP16 quality. For 12+ GB cards in GGUF workflows.
Q4_0 GGUF (around 2.7 GB) — 4-bit quantized. Smallest file. Makes SDXL actually run on 6–8 GB entry-level cards.
Quants Explained — Which File Do I Pick?
If you've ever seen FP16, BF16, FP8, Q4, Q8 and just downloaded the biggest one, this section is for you.
What's a quant? Same model, smaller file. Weights are compressed so they fit on less VRAM. Some quality loss vs FP16, but smart compression (Q8_0) is so close you won't see a difference in normal use.
Pick by Your VRAM
24 GB+ (3090, 4090, 5090, A6000) — FP16 or BF16 from my other post. No reason to compress.
12–16 GB (3060 12GB, 4070, 4080) — FP8 Scaled or Q8_0 GGUF. Near-FP16 quality with headroom for LoRAs.
8–12 GB (3060, 4060 Ti, 2080) — Q8_0 GGUF. Solid quality, comfortable VRAM use.
6–8 GB (3050, 2060, 1660) — Q4_0 GGUF. Smallest file, makes SDXL actually work.
CPU only or 4 GB cards — Q4_0 GGUF in ComfyUI-GGUF. Slow but functional.
Quality Ladder
FP16 ≈ BF16 ≈ Q8_0 > FP8 > Q4_0. Above Q4_0 the differences are basically invisible in normal generation.
About Speed
Smaller quants are NOT always faster. Generation speed is mostly compute-bound on most cards — quants help with VRAM fit, not raw iterations per second. Where they DO help speed: avoiding system-RAM offload, which is what kills speed on small cards when the model doesn't fit.
DMD2 Usage
DMD2 is a different beast. It's the FP16 model with a distillation LoRA merged in, so it generates in 4 steps instead of 25–30.
Sampler — LCM
Steps — 4 (not 25+)
CFG — 1.2 (not 7)
Result — Comparable quality to a 25-step generation in roughly 1/6 the time
Works on any card that fits SDXL. Great for batch work or fast iteration.
How to Load GGUF Files
GGUFs need a loader, since most UIs don't natively support them yet.
For ComfyUI — install the ComfyUI-GGUF custom node:
https://github.com/city96/ComfyUI-GGUF
For Forge or Forge Neo — install my Forge SDXL GGUF extension:
https://github.com/brandulateai/sd-forge-sdxl-gguf-brandulateai
After installing, GGUFs load straight from the standard checkpoint dropdown. No external module picker, no extra setup.
All my GGUFs are bundled (UNet + CLIP-L + CLIP-G + VAE in one file) so they load without picking separate components.
Trigger Word
99rbsy99 — same as the main release. Add to every prompt for V6.5 realism. Stacks cleanly with character LoRAs using 99bsy99.
Pro Version Available
This is the standard version of V6.5. The Pro version is trained on more data for longer, producing a more polished and refined output. Get the Pro version of my models here
Found Anything Off?
Drop it in the comments or on Discord. V7's fix list starts now share your feedback here.
FAQ
Comments (13)
The Q_8 GGUF version of 6.5-DMD2 does not seem to work correctly (it produces output similar to a non-DMD model at low steps). The safetensors version works as expected.
Q8 and Q4 are not DMD2. They are normal Models.
@Stable_Yogi Then why is the filename realismByStableYogi_ponyV65DMD2Gguf.gguf ?
@psammites Civitai adds that at the end, based on the model card page., look at the model variant section, it mentions all models and their properties.
Hi,
Tks for your works.
I tried it but it always give me a "doll" impression, not that much realistics and fingers are usualy bad.
I tried the 12GB et 6GB model with your lora end embedings
Steps 27 - is it best, or 30 would be cool too?
Try between 18 and 36.
27 seems to be the mid ground.
nice model
goodmd
❤️
❤️
It doesn't seem like you or your contributors are using the positive/negative embeddings any more. Is that just for the older models?
Good eye. It's mostly an older-model thing. The newer trained ones handle anatomy on their own now, so they need a lot less hand-holding and I barely reach for the anatomy embeddings with them. The realism and general negatives I still lean on though, those stay effective no matter the model. Newer builds are all up on my portal if you wanna grab them. Hope that clears it up.