Krea-Char-Design-Data

title: Krea 2 Character Design
emoji: 🎨
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: other
short_description: Krea 2 Character Design LoRA on Krea 2 Turbo (ZeroGPU-ready)

🎨 Krea 2 · Character Design

A Gradio Space that runs the CivitAI [KREA 2] Character Design LoRA on top of krea/Krea-2-Turbo using the diffusers Krea2Pipeline.

It generates proper character design sheets — front / side / back views, expression studies, colour palettes and accessories.

How it works

The LoRA lives in the companion HF dataset (models/lora/krea2-character-design.safetensors) and is downloaded into a local models/lora/ folder on startup.
Inference runs inside a @spaces.GPU allocation (ZeroGPU-ready).
Every generation (image + prompt + settings) is archived back to the dataset under generations/.

⚙️ Required setup

This Space depends on gated models and a write-enabled token:

Accept the Krea 2 license with the account whose token you use:
- https://huggingface.co/krea/Krea-2-Turbo (and optionally krea/Krea-2-Raw)
Add a Space secret named HF_TOKEN = a HF write token (needed for the gated base model and to save generations to the dataset).
Hardware: currently set to CPU. Switch to ZeroGPU (Settings → Hardware) to actually run the 12B model — it will not generate on CPU.

Environment variables (optional)

Var	Default
`BASE_MODEL`	`krea/Krea-2-Turbo`
`DATASET_REPO`	`that-username-is-not-available/Krea-Char-Design-Data`
`LORA_REPO_PATH`	`models/lora/krea2-character-design.safetensors`
`SAVE_GENERATIONS`	`1`

Recommended settings (Turbo)

Steps: 8 · Guidance: 0.0 · LoRA strength: 1.0
Trigger word Character design is auto-prepended (toggle in the UI).