FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition

Read the paper on arXiv: FlowSlider Paper

FlowSlider lets you control how much an image edit happens—from subtle changes to dramatic transformations.

How It Works

The magic is in separating the edit dynamics into two independent parts:

  • Fidelity — keeps your image looking like the original
  • Steering — pushes the image toward your target description

By adjusting the strength slider s, you can amplify the steering effect while keeping the fidelity anchor intact, giving you smooth continuous control over the edit intensity.

Try it: Upload an image, describe what you see and what you want to change, then slide to find your perfect level of edit intensity!

Examples

Each strip shows the original image followed by FlowSlider outputs at strengths s = 1 → 5.

Decay: Metal mugs — Add rust, corrosion, damage, and overgrowth to metal mugs

Season: Summer → Winter — Change the season to winter with snow

Season: Autumn → Spring — Change the season to spring with fresh green leaves

Season: Spring → Autumn — Change the season to autumn with warm fall colors

Time of Day: Overcast → Sunset — Change the time to sunset with golden light

Try It Yourself

⚠️ Note: Due to HuggingFace Spaces resource limits, results are resized to 512px on the short edge and may take ~30 seconds to generate.
Backbone Model

FLUX.1-dev is recommended. Switching reloads model weights on first use.

Prompts
10 100
1 60
1 10
1 20
Load an example
Backbone Model Source Image Source Prompt (describe the original image) Target Prompt (describe the desired edit) Negative Target Prompt (optional) Edit Strengths (s) T steps n_max Source Guidance Scale Target Guidance Scale Seed

Paper: FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition Backbones: FLUX.1-dev · Stable Diffusion 3