Model Hub & GGUF Quantization Presets
Specifications for built-in mobile-optimized presets and custom weight resolution.
| Preset Name | Base Architecture & Quantization | File Size | Optimal Steps & CFG | Recommended Sampler | Key Visual Workload |
|---|---|---|---|---|---|
"sdxs" |
SDXS 512 Tiny SD Distilled (Q8_0) | 651 MB | 1 ~ 2 steps (CFG 1.0) | euler_a |
Ultra-low latency mobile prototyping (Instant 1-2s, sharp) |
"anime" |
DreamShaper 8 LCM (Q4_0) | 1.55 GB | 4 ~ 8 steps (CFG 1.5) | lcm |
2D / 2.5D stylized anime art (Crisp lineart, rich cel-shading) |
"realistic" |
Realistic Vision V6.0 B1 (Q4_K) | 1.55 GB | 20 ~ 25 steps (CFG 7.0) | dpm2 / karras |
Ultra-detailed photorealism (Pores, eyes, cinematic lighting) |
"speed" |
Stable Diffusion 1.5 Base (Q4_1) | 1.68 GB | 15 ~ 20 steps (CFG 6.0) | euler_a / dpm++2m |
General-purpose drafting and balanced composition |
"turbo" |
Stable Diffusion 1.5 Pruned (Q4_0) | 1.49 GB | 15 ~ 20 steps (CFG 6.0) | euler_a / dpm++2m |
Lightweight SD1.5 base generation |
đĄ Denoising Architecture Rules (ė ė íëŧë¯¸í° ę°ė´ë)
1. Distilled Models (sdxs, anime): Keep CFG low (1.0~1.5) and use 1st-order samplers (euler_a, lcm). High CFG or 2nd-order ODE samplers (dpm2) will collapse the latent space.
2. Full SD1.5 Models (realistic, speed, turbo): Require at least 15~20 steps and CFG 6.0~7.5 with quality-guard negative prompts to fully resolve photorealistic details.
Custom Model Management API
from termux_diffusion import (
set_cache_dir, # Route cache to external storage / SD card
get_cache_dir, # Inspect active cache directory
download_model, # Pre-download models in background with progress
register_model, # Register custom Hugging Face GGUF models
list_cached_models, # Inspect downloaded models
clear_cache # Purge cache to reclaim storage
)
# 1. Configure custom cache storage path (e.g. SD Card)
set_cache_dir("~/storage/external-1/ai_models")
# 2. Pre-fetch preset weights
download_model("sdxs", force=False)
# 3. Register custom repository alias
register_model(
name="waifu",
repo_id="second-state/DreamShaper-8-GGUF",
filename="dreamshaper-8-Q4_k.gguf",
description="DreamShaper 8 Q4_K model for stylized anime portraits"
)
# 4. View cached weights
models = list_cached_models()
for m in models:
print(f"Model: {m['name']}, Size: {m['size_mb']:.1f}MB")