v1-5-pruned-emaonly-fp16.safetensors
🔍 Overview
The v1-5-pruned-emaonly-fp16.safetensors checkpoint is a Stable Diffusion v1.5-based model, optimized for inference with half-precision (FP16) weights. It's a pruned, Exponential Moving Average (EMA)-only version, stripped of training overhead and tuned for speed and efficiency in rendering high-quality text-to-image generations.
If you're looking for a rock-solid, vanilla foundation for anything from photorealism to stylized illustration—and you enjoy buttery-smooth memory performance—this checkpoint should be your default starting block.
📁 Technical Details
| Property | Value |
|---|---|
| Base Model | Stable Diffusion v1.5 |
| Architecture | UNet + CLIP text encoder |
| Precision | FP16 (half-precision) |
| Pruning | Yes (non-essential weights removed) |
| EMA Only | Yes (uses Exponential Moving Average weights) |
| File Format | .safetensors (secure and efficient) |
| Size | ~2.1GB |
| CLIP Version | OpenCLIP ViT-H/14 (compatible with SD 1.5 text encoders) |
🧠 What It Can Do
This checkpoint serves as a general-purpose generation model and performs well across a range of prompt styles including:
- 🧍 Realistic portraits and character renders
- 🌄 Landscapes and environmental scenery
- 🎨 Artistic interpretations and semi-abstracts
- 🧙 Fantasy and concept art
- 🧸 Stylized illustration (when paired with appropriate LoRAs/VAEs)
🧰 Recommended Use Cases
This model is particularly useful for:
- Baseline workflows: It’s the foundation model often used to compare LoRAs, VAEs, and prompt tweaking.
- Resource-constrained environments: Because it’s FP16 and pruned, it plays nicely with low VRAM (yes, 4–6GB cards can actually have fun here).
- Compositional consistency testing: With proper CFG control and noise seeds, this model handles consistent framing well.
- Fine-tuning or LoRA training: It's often used as a base checkpoint for finetuning projects due to its stability.
🧩 Compatible VAEs
This checkpoint doesn’t include a baked-in VAE. You’ll want to pair it with an external .safetensors VAE file. Here are the top contenders:
| VAE Name | Recommended Use |
|---|---|
vae-ft-mse-840000-ema-pruned.safetensors | Clean, accurate, default VAE for SD 1.5 |
kl-f8-anime2.ckpt | For more stylized/illustrative outputs |
clearVAE or colorfulVAE | Boost contrast and vibrancy |
If you're getting muddy or washed-out images, check your VAE. Misalignments here can ruin even the best prompts.
⚙️ ComfyUI Settings & Parameters
| Parameter | Suggested Value | Notes |
|---|---|---|
| Sampler | Euler a, DPM++ 2M Karras, UniPC | Euler a = faster, DPM++ = better quality |
| Steps | 20–35 | Sweet spot is often 28 |
| CFG Scale | 6.5–8.5 | Over 9 can create artifacts or melting faces |
| Resolution | 512x512 native, upscale from there | Upscaling recommended post-gen |
| Seed | Fixed for reproducibility | Variants = use dynamic seed |
💬 Prompting Tips
- Stick to natural language prompts with clear modifiers. Example:
arduino
CopyEdit
"A futuristic cityscape at sunset, cyberpunk aesthetic, neon lights, volumetric fog, ultra detailed, unreal engine, 8k" - Add specific modifiers to control mood and fidelity:
ultra-detailed,intricate lighting,masterpiece,high quality
- Negative prompts are essential to avoid those "why does she have six fingers and a third eye?" moments:
scss
CopyEdit
(worst quality, low quality:1.3), bad anatomy, deformed, blurry, extra limbs, mutated hands, bad proportions
🚨 Known Issues & Limitations
| Issue | Details |
|---|---|
| Lack of baked-in VAE | You must attach a compatible VAE node in ComfyUI or you’ll get low-fidelity outputs. |
| Limited stylization | It’s not anime or toon-optimized out of the box—pair with a LoRA for stylization. |
| Occasionally bland outputs | When prompting is too vague, it plays safe. So don’t be boring. |
| Small resolution default | Works best at 512x512; higher resolutions may need tiled workflows or upscaling nodes. |
🧪 Ideal Pairings
- LoRAs: Combine with LoRAs that add style or fine-grained control (e.g.,
analog-style,cyberpunk-city,gothicFashion_v1.0). - Upscalers:
4x-UltraSharp,R-ESRGAN 4x+, or ComfyUI’sUltimateSDUpscalefor crisp final images. - Post-processing: Use
Prompt Styler,Hires Fix, orDetail Tweakernodes to push visual fidelity even further.
📚 Additional Resources
🧠 Final Thoughts
The v1-5-pruned-emaonly-fp16.safetensors checkpoint is your Swiss Army knife in the ComfyUI arsenal. It might not breathe fire out of the box like some hyper-stylized blends, but it’s damn reliable, exceptionally versatile, and light on resources.
Use it to test, to fine-tune, to baseline, or as your dependable workhorse for clean, quality generations.
Want model magic? Start here—just don't forget your VAE, or you’ll be asking why everything looks like it was shot through a potato.