VisualNote Extra Detailer XL (Style, Art Texture)

LORA
Original


Updated:

DPM++ samplers are recommended.

🔹 Karras vs Normal (Scheduler)

Normal (linear/exponential) = evenly distributes timesteps. Good enough, but less efficient.

Karras = clusters more steps near the end (low-noise region) → spends more effort refining tiny details → better for grain, freckles, pores, brush textures.

👉 Most people pair DPM++ 2M/3M SDE + Karras for realism + detail retention.

👉 If you want more “looseness” or organic look, use DPM++ 2S a (Karras) or even Euler a.

LoRA

Extra Detailer XL

strength:

0.1 (recommended)

0.2-0.6 (ok)

0.7 (experimental)

0.8-1.0 (Experimental Warning: over saturation, over highlighting, over taking styles, reducing/removing other LoRA styles and texture effects)

Recommended CFG:

Euler/Euler a:

3-7

DPM/DPM++ samplers:

3

2.5 - 3.5 (experimental)

2 - 4 (more experimental)

5 - ... (Experimental Warning: causes over saturation, over exaggerated weird details, over highlighting)

Euler/Euler a - good for Simple, Flat anime/illust styles but

Flattens/Removes Fine-Details

like art textures, textures for film effects, visual effects, traditional art medium, dust, fabric, hair, etc.

1.: Strengths

2.: Weaknesses

Euler a

Simple + ancestral noise

1. Fast, stylized, organic variation

2. Eats fine texture, less sharp, inconsistent

DPM2 a

2nd-order + noise

1. Sharper than Euler a, some variation

2. Outdated vs DPM++

DPM++ 2S a

2nd-order single-step + noise

1. Punchy, organic, slightly sharper than Euler a

2. Less stable at low steps

DPM++ 2M SDE

2nd-order multi-step SDE

1. Stable, detailed, great texture

2. Slightly slower, less random

DPM++ 3M SDE

3rd-order multi-step SDE

1. Maximum sharpness, photorealism, ultra-detail

2. Rigid, slower, less stylistic flexibility

Version Detail

Trained by Tensor
SDXL 1.0
1
LoRA Focus: • Balanced across high-quality and diverse content, with: • Wide range of unique art styles & distinct characters • Strong emphasis on sketch, pencil, and illustration work • Inclusion of historical, traditional, and tribal aesthetics • Classic/retro influences from 1980s–2000s • Subtle touches of cartoon, comic, and semi-realism • Cleaned complex, detailed backgrounds for clarity • Cleaning of old/low-quality images, removing over-exaggerated colors or textures Training LORA details: Sampler: dpmpp_3m_sde_gpu: this is very good for detail preservation. 🛠️ Optional Improvements for More Control Use cosine LR scheduler with warmup: improves learning for longer training. lr_scheduler=cosine_with_restarts, num_warmup_steps=100 Learning Rates Unet LR: 0.0001 is generally good for style LoRAs. Text Encoder LR: 0.00001 is good for subtle style influence. clip_skip = 2 is often better for anime models (like NAI, Pony, NoobAI, etc.). Min-SNR Gamma using 5, which is ideal for preserving features in high-quality datasets. This is recommended for current SDXL/SD1.5 workflows. Network Dim and Alpha Dim (default 64): Higher values (e.g. 128) capture more detail but increase file size and may overfit. Alpha (default 32): Should be around 0.5×Dim. Keep as-is unless experimenting. ✅ Higher dim usually = better quality, but more VRAM needed. 2025-08-07 21:52:20

Project Permissions

Reprinting is strictly prohibited

    Use Permissions

  • Use in TENSOR Online

  • As a online training base model on TENSOR

  • Use without crediting me

  • Share merges of this model

  • Use different permissions on merges

    Commercial Use

  • Sell generated contents

  • Use on generation services

  • Sell this model or merges

Related Posts