README.md

metadata

title: Krea 2 Ostris Edit
emoji: 🎨
colorFrom: pink
colorTo: green
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
pinned: false
hardware: zero-a10g
short_description: Reference-image conditioned generation with Krea 2
python_version: '3.10'
startup_duration_timeout: 30m
models:
  - krea/Krea-2-Turbo
  - ostris/Krea2OstrisEdit
  - ostris/krea2_turbo_style_reference

Krea 2 Ostris Edit

Reference-image conditioned image generation with Krea 2 Turbo, powered by the Krea2OstrisEdit community pipeline.

How it works

Upload a reference image and type a prompt describing what you want to generate. The Style Reference LoRA conditions the Krea 2 Turbo model in two ways:

1. Text encoder — the reference image is embedded into the prompt via Qwen3-VL vision placeholders, so the text conditioning "sees" the reference
2. Latent space — clean VAE latents of the reference are appended to the transformer sequence at flow time t=0 (Kontext-style "index" placement)

Toggle the "Use reference image" checkbox to switch between reference-conditioned generation and plain text-to-image mode.

Components

• Base model: krea/Krea-2-Turbo (8-step distilled Krea 2)
• Custom pipeline: ostris/Krea2OstrisEdit (reference-image conditioning + AI-Toolkit LoRA loading)
• LoRA: ostris/krea2_turbo_style_reference (style reference conditioning)

Inference recipe

• Steps: 8 (Turbo distilled)
• Guidance: 0.0 (disabled, per Krea 2 Turbo convention)
• LoRA scale: 1.0 (adjustable)
• Resolution: 1024x1024 (adjustable, must be multiples of 16)