Sign In

LucentXL Pony by klaabu

864

4.5k

296.9k

285

Verified:

SafeTensor

Type

Checkpoint Trained

Stats

304

2.5k

Generation

Reviews

Published

Jul 3, 2025

Base Model

Pony

Hash

AutoV2
7A67CAFAC4
GPU (Great Purring Unit)
klaabu's Avatar

klaabu

b2.0

New Release: LucentXL (untitled . pony Fine-Tuned Upgrade)

The original untitled.pony laid the foundation. This version builds on it β€” deeper, smarter, sharper. LucentXL is a fine-tuned and expanded upgrade, refined for improved photorealism and generation fidelity across a wider range of prompts and styles.

πŸ”§ Key Improvements:

  • New name, same roots – renamed based on community feedback, while keeping lineage intact.

  • Fine-tuning on enhanced dataset – expanded with additional photorealistic and ethnic-specific material for broader diversity and realism.

  • Improved sampler compatibility – tuned to perform exceptionally with DPM++ 2M Karras, giving more stable and natural output with less tweaking.

  • Increased token comprehension – upgraded text encoder awareness and prompt reactivity.

  • Ethnicity-aware – better representation and understanding of diverse facial features and cultural cues.

  • Precision detailing – finer control over eyes, lighting, skin texture and tone.

  • Block weight adjustments – internal tweaks to balance realism layers without oversaturation or hallucinations.


πŸ”§ Settings

Local Generation

  • Sampler: DPM++ SDE

  • CFG: 4.5

  • Steps: 20+

On Civit

  • Sampler: DPM++ 2M Karras

  • CFG: 4.5

  • Steps: 25+

πŸ”Upscaling (Latent)

  • Method: Latent

  • Resize: 1.5x

  • Denoise strength: 0.5

  • Steps: 15

πŸ“ Latent upscaling benefits from the model’s own internal fine-tuned understanding of lighting, skin texture, and soft anatomical detail, giving a natural enhancement without introducing artefacts or plastic smoothing.

‐-----------------------------------------------------------------------------

b1.0

Untitled is an experimental checkpoint built for users who want better control, clarity, and prompt consistency across stylised and semi-realistic generations. It's flexible, designed for creators who know what they want and want it rendered clean. It also features enhanced token adherence, improving its ability to accurately interpret and respond to a wider range of prompts

This checkpoint performs significantly better in local environments due to sampler and backend limitations on Civitai. Prompt interpretation, lighting response, and shape consistency all benefit from proper sampler use.


βœ… On Civitai (Web UI):

  • Sampler: DPM++ 2M Karras

  • CFG: 4 – 6

  • Steps: 20 – 35

βœ… Local Use (e.g. A1111, ComfyUI):

  • Sampler: DPM++ SDE

  • CFG: 2 - 5

  • Steps: 20 – 35

  • Works well with ControlNet, IPAdapter, LoRAs, and realism-focused workflows.


Still experimental. Expect occasional variation β€” but when it hits, it hits clean.
Feedback and test images welcome.