Available now

Stable Diffusion with LoRAs: inpainting

Stable Diffusion with LoRAs runs a Stable Diffusion base model you choose, with your own LoRA weights, ControlNets and IP adapters, from text, from an image or with a mask.

Model ID

Input

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

Link or model ID of the base model.

Be as descriptive as possible for best results.

Example prompt. Edit it or write your own.

The image to use for image to image/inpainting.

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

The black-and-white image to use as mask during inpainting.

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

Make sure to use a background compatible with the model.

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

The IC Light model image to use for the image generation.

Uploading a file needs an account. Create account

JPEG, PNG, WebP · up to 25 MB

Number of images to generate in one request. Note that the higher the batch size, the longer it will take to generate the images.

Additional settings

Link or model ID of a custom U-Net model.

The model variant to use, for example fp16.

If set to true, the prompt weighting syntax will be used. Additionally, this will lift the 77 token limit by averaging embeddings.

The amount of noise to add to noise image for image. 1.0 is complete noise and values near 0 are minimal noise.

You can use up to 10 LoRAs and they will be merged together to generate the final image.

Only a single embedding is supported at the moment. The embeddings will be used to map the tokens in the prompt to the embedding weights.

If set to true, the controlnet will be applied to only the conditional predictions.

IP adapters for image generation (up to 5).

Choose an encoder compatible with the IP-Adapter weights.

The weight name of the image encoder model to use for the image generation.

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

Skips part of the image generation process, leading to slightly different results. This means the image renders faster, too.

Scheduler / sampler to use for the image denoising process.

The epsilon is the default.

The size of the tiles to be used for the image generation.

The size of the tiles to be used for the image generation.

The eta value to be used for the image generation.

Some published inputs are not in this form.

  • Timesteps: Nested object inputs are not supported yet. It is not sent.
  • Sigmas: Nested object inputs are not supported yet. It is not sent.
Price$0.60 / run

You sign up first. Your settings are kept.

About this model

Overview

This card is for people who already work with Stable Diffusion files. You give a link to the base model and add as many LoRAs as you like, each with a scale from 0 to 4; they are merged into one result. ControlNets, one embedding and IP adapters can be added the same way.

Three modes share these settings: text to image, image to image with a noise strength, and inpainting with a black-and-white mask. You also choose the scheduler, steps up to 150, guidance scale and clip skip.

What you can build

  • Images in a character or style you trained as a LoRA
  • Poses or outlines held in place by a ControlNet
  • Several LoRAs blended at different strengths
  • Inpainting fixes with a community checkpoint

Prompting tips

  • Start each LoRA at a scale of 1 and adjust one at a time, so you can see what each one changes.
  • Use the trigger words of your LoRA or embedding in the prompt; without them the style may not show.
  • In image to image, a noise strength near 1 replaces almost everything; stay near the middle to keep the layout.

What to check

Check subject consistency and unwanted style spillover; inspect repaired edges when using inpaint.

Inputs

  • Model Name (required): Link or model ID of the base model.
  • Unet Name (optional): Link or model ID of a custom U-Net model.
  • Variant (optional): The model variant to use, for example fp16.
  • Prompt (required): The prompt to use for generating the image. Be as descriptive as possible for best results.
  • Negative Prompt (optional): The negative prompt to use.Use it to address details that you don't want in the image. This could be colors, objects, scenery and even the small details (e.g. moustache, blurry, low resolution).
  • Prompt Weighting (optional): If set to true, the prompt weighting syntax will be used. Additionally, this will lift the 77 token limit by averaging embeddings.
  • Image (optional): URL of image to use for image to image/inpainting.
  • Mask (optional): URL of black-and-white image to use as mask during inpainting.
  • Noise Strength (optional): The amount of noise to add to noise image for image. 1.0 is complete noise and values near 0 are minimal noise.
  • Loras (optional): The LoRAs to use for the image generation. You can use up to 10 LoRAs and they will be merged together to generate the final image.
  • Embeddings (optional): The embeddings to use for the image generation. Only a single embedding is supported at the moment. The embeddings will be used to map the tokens in the prompt to the embedding weights.
  • Controlnets (optional): The control nets to use for the image generation. You can use up to 5 control nets and they will be applied to the image at the specified timesteps.
  • Controlnet Guess Mode (optional): If set to true, the controlnet will be applied to only the conditional predictions.
  • Ip Adapter (optional): IP adapters for image generation (up to 5).
  • Image Encoder Path (optional): Choose an encoder compatible with the IP-Adapter weights.
  • Image Encoder Subfolder (optional)
  • Image Encoder Weight Name (optional): The weight name of the image encoder model to use for the image generation.
  • Ic Light Model URL (optional): The URL of the IC Light model to use for the image generation.
  • Ic Light Model Background Image (optional): The URL of the IC Light model background image to use for the image generation. Make sure to use a background compatible with the model.
  • Ic Light Image (optional): The URL of the IC Light model image to use for the image generation.
  • Seed (optional): The same seed and the same prompt given to the same version of Stable Diffusion will output the same image every time.
  • Number of inference steps (optional): Increasing the amount of steps tells Stable Diffusion that it should take more steps to generate your final result which can increase the amount of detail in your image. From 1 to 150.
  • 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.
  • Clip Skip (optional): Skips part of the image generation process, leading to slightly different results. This means the image renders faster, too. From 0 to 2.
  • Scheduler (optional): Scheduler / sampler to use for the image denoising process. Options: DPM++ 2M, DPM++ 2M Karras, DPM++ 2M SDE, DPM++ 2M SDE Karras, Euler, Euler A, Euler (trailing timesteps), LCM, LCM (trailing timesteps), Ddim, TCD.
  • Prediction Type (optional): The type of prediction to use for the image generation. The `epsilon` is the default. Options: V prediction, epsilon.
  • Rescale Betas Snr Zero (optional)
  • Image Format (optional): The format of the generated image. Options: jpeg, png.
  • Number of images (optional): Number of images to generate in one request. Note that the higher the batch size, the longer it will take to generate the images. From 1 to 8.
  • Tile Width (optional): The size of the tiles to be used for the image generation. From 128 to 4096.
  • Tile Height (optional): The size of the tiles to be used for the image generation. From 128 to 4096.
  • Tile Stride Width (optional): The stride of the tiles to be used for the image generation. From 64 to 2048.
  • Tile Stride Height (optional): The stride of the tiles to be used for the image generation. From 64 to 2048.
  • Eta (optional): The eta value to be used for the image generation. From 0 to 1.

Output and limits

  • Up to 8 images per run.
  • You need links to the base model and to each LoRA, ControlNet or adapter file; the card does not include a library of them.
  • 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