"if you want other characters (maybe less difficult than Ellie :)) just say so in the comments"
Ellie_TLOU_XL is an identity‑focused SDXL LoRA trained on approximately 80 original images, all created exclusively from scratch for this dataset. No downloaded material or external images were used. Every image was generated, selected, and refined manually to ensure full control over identity, lighting, and consistency.
The dataset includes:
high‑quality portraits
half‑body shots
neutral and environmental lighting
narrative scenes
recognizable locations inspired by the videogame’s world, enhancing atmosphere and contextual coherence
Training was performed on SDXL and intentionally stopped at 1750 steps, the optimal point before overfitting. This preserves:
natural skin texture
stable proportions
strong identity consistency
zero artifacts
reliable performance even under complex lighting
Best Checkpoint: The model performs exceptionally well with SDXL Base, achieving maximum identity fidelity and the most natural rendering. It also works with other neutral checkpoints, but SDXL Base remains the recommended choice.
Recommended Strength:
0.75 → maximum stability; ideal for half‑body, narrative scenes, and complex environments
0.85 → professional sweet spot; strong identity with natural skin
0.9 → maximum safe strength; extremely faithful identity without introducing artifacts
Ideal Use Cases:
Excellent for portraits (highest likeness, clean skin, precise details)
Very good for half‑body shots (stable proportions, consistent identity)
Reliable in narrative scenes (environment, dramatic lighting, post‑apocalyptic mood)
Description
⭐ Ellie – Z‑Image Turbo Edition
Optimized version for Z‑Image Turbo (ZIT)
This LoRA was trained natively on Z‑Image Turbo, not on SDXL. Because of that, it behaves differently from the SDXL version:
faster response
stronger detail retention
more cinematic rendering
stable skin texture and micro‑details
better handling of complex lighting
improved consistency in half‑body and narrative scenes
⭐ Recommended Weight Range
✔ 0.65 – 0.90 → Optimal Z‑Image Turbo zone
In this range the LoRA performs exactly as intended:
stable identity
natural skin
coherent eyes
no artifacts
cinematic look
✔ For a more realistic look
👉 Lower the weight (0.45 – 0.60) ZIT takes the lead and the LoRA becomes a realism enhancer.
✔ For a more “videogame / concept art” look
👉 Increase the weight (0.80 – 0.95) Identity becomes stronger, more stylized, more game‑ready.
⭐ Trigger Word
Use:
ellie_tlou
Short, clean, non‑intrusive, and activates the identity naturally.
⭐ Why this version differs from SDXL
The SDXL version is softer, more neutral, more photographic. The Z‑Image Turbo version is:
sharper in micro‑details
more consistent in close‑ups
better at preserving skin texture
more stable under dramatic lighting
stronger in half‑body compositions
ideal for narrative scenes (gas mask, gear, exploration)
It’s a more modern, cinematic, and robust version.