Available now

FLUX.1 Differential Diffusion: image editing

By Black Forest Labs

FLUX.1 Differential Diffusion edits an image area by area: a change map sets how much each part may change while your prompt describes the result.

Model ID
Maker
Black Forest Labs

Input

Fields marked * are required. Review the price before generating.

Example prompt. Edit it or write your own.

This is always set to 1 for streaming output.

The Image for Reference-Only

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

The image to use as initial image.

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

Additional settings

The number of inference steps to perform.

The same seed and the same prompt given to the same version of the model will output the same image every time.

You can use any number of LoRAs and they will be merged together to generate the final image.

You can use any number of LoRAs and they will be merged together to generate the final image.

Only one controlnet is supported at the moment.

Only one controlnet is supported at the moment.

Use an image input to influence the generation. Can be used to fill images in masked areas.

The images to be filled for redux prompting

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

Uses classical CFG. Generation takes longer when this is on.

Uses CFG-zero init sampling.

Only used if a reference image is provided.

The percentage of the total timesteps when the reference guidance is to bestarted.

The percentage of the total timesteps when the reference guidance is to be ended.

Base shift for the scheduled timesteps

Max shift for the scheduled timesteps

Specifies whether beta sigmas ought to be used.

Sigmas schedule for the denoising process.

Scheduler for the denoising process.

Negative prompt to steer the image generation away from unwanted features.

Higher values will result in a image that is more distant to the negative prompt.

After the specified proportion of steps has been iterated, the remaining steps will use original attention processors in FLUX.

The strength to use for differential diffusion. 1.0 is completely remakes the image while 0.0 preserves the original.

Price$0.60 / run

You sign up first. Your settings are kept.

About this model

Overview

Upload the image, a change map and a prompt. The shades of the map give each region its own amount of change, so one part can be redrawn while another stays close to the original, with soft transitions in between. Strength sets the overall amount.

It runs on FLUX.1 [dev] and accepts LoRAs, a ControlNet, IP-Adapters and a reference image, plus a negative prompt. Each run returns one to ten images as PNG or JPEG.

What you can build

  • A new sky or season while the foreground stays as it is
  • Product shots with a changed background and an untouched product
  • Gradual blends between an edited area and the original photo
  • Style edits limited to one part of a picture

Prompting tips

  • Use soft gradients in the change map where edited and kept areas meet, so the change fades in gradually.
  • Write the prompt as a description of the whole finished image, not only the part that changes.
  • Lower the strength when the changed area drifts too far from the original.

What to check

Inspect transitions around edited regions and compare areas meant to retain their original appearance.

Inputs

  • Prompt (required): The prompt to generate an image from.
  • Image Size (optional): The size of the generated image. Options: Square HD, Square, Portrait 4:3, Portrait 16:9, Landscape 4:3, Landscape 16:9.
  • Number of Inference Steps (optional): The number of inference steps to perform. From 1 to 50.
  • Seed (optional): The same seed and the same prompt given to the same version of the model will output the same image every time.
  • Loras (optional): The LoRAs to use for the image generation. You can use any number of LoRAs and they will be merged together to generate the final image.
  • Control Loras (optional): The LoRAs to use for the image generation which use a control image. You can use any number of LoRAs and they will be merged together to generate the final image.
  • Controlnets (optional): The controlnets to use for the image generation. Only one controlnet is supported at the moment.
  • Controlnet Unions (optional): The controlnet unions to use for the image generation. Only one controlnet is supported at the moment.
  • Ip Adapters (optional): IP-Adapter to use for image generation.
  • Easycontrols (optional): EasyControl Inputs to use for image generation.
  • Image Fill (optional): Use an image input to influence the generation. Can be used to fill images in masked areas.
  • Guidance scale (CFG) (optional): The CFG (Classifier Free Guidance) scale is a measure of how close you want the model to stick to your prompt when looking for a related image to show you. From 0 to 20.
  • Real CFG scale (optional): The CFG (Classifier Free Guidance) scale is a measure of how close you want the model to stick to your prompt when looking for a related image to show you. From 0 to 5.
  • Use Real CFG (optional): Uses classical CFG. Generation takes longer when this is on.
  • Use CFG-Zero-Init (optional): Uses CFG-zero init sampling.
  • Number of Images (optional): The number of images to generate. This is always set to 1 for streaming output. From 1 to 10.
  • Reference Image (optional): URL of Image for Reference-Only.
  • Reference Strength (optional): Only used if a reference image is provided. From -3 to 3.
  • Reference Start (optional): The percentage of the total timesteps when the reference guidance is to bestarted. From 0 to 1.
  • Reference End (optional): The percentage of the total timesteps when the reference guidance is to be ended. From 0 to 1.
  • Base Shift (optional): Base shift for the scheduled timesteps. From 0.01 to 5.
  • Max Shift (optional): Max shift for the scheduled timesteps. From 0.01 to 5.
  • Output Format (optional): The format of the generated image. Options: jpeg, png.
  • Use Beta Schedule (optional): Specifies whether beta sigmas ought to be used.
  • Sigma Schedule (optional): Sigmas schedule for the denoising process.
  • Scheduler (optional): Scheduler for the denoising process. Options: euler, Dpmpp 2m.
  • Negative Prompt (optional): Negative prompt to steer the image generation away from unwanted features.
  • NAG scale (optional): The scale for NAG. Higher values will result in a image that is more distant to the negative prompt.
  • NAG Tau (optional): The tau for NAG. Controls the normalization of the hidden state. Higher values will result in a less aggressive normalization, but may also lead to unexpected changes with respect to the original image. Not recommended to change this value.
  • NAG alpha (optional): The alpha value for NAG. This value is used as a final weighting factor for steering the normalized guidance (positive and negative prompts) in the direction of the positive prompt. Higher values will result in less steering on the normalized guidance where lower values will result in considering the positive prompt guidance more.
  • Proportion of steps to apply NAG (optional): The proportion of steps to apply NAG. After the specified proportion of steps has been iterated, the remaining steps will use original attention processors in FLUX.
  • Image (required): URL of image to use as initial image.
  • Change Map URL (required): URL of change map.
  • Strength (optional): The strength to use for differential diffusion. 1.0 is completely remakes the image while 0.0 preserves the original.

Output and limits

  • Image sizes: Square HD (1024×1024), Square (512×512), Portrait 4:3 (768×1024), Portrait 16:9 (576×1024), Landscape 4:3 (1024×768), Landscape 16:9 (1024×576).
  • File formats: JPEG, PNG.
  • Up to 10 images per run.
  • Small or dense text in the image can come out misspelled.
  • Hands, crowds and complex poses can show artifacts; generate a few variations and pick the best.
  • Some prompts or outputs can be refused by safety checks.

Price guide

  • up to: $0.60 / run