This page describes the basic operation of Img2Img. If you are confused about the parameters, please see the WebUi parameters on the Model Debugging page.
If you are looking for some application examples, you should see the "Practical Guide" page.
Please read the previous section, Model Advance 1, for specific Img2Img and inpaint introduction operations.
cfg scale corresponds to the degree of prompt, the higher the value the more literal the prompt will be, the lower it will give the model more room to play, but the actual model performance in terms of CFG scale low (6-8) is low saturation, on the cluttered side, high (18-22) is high saturation, more intense style.
High CFG will cause color distortion, CFG should be between 5-15
denoise strength img2img proprietary parameter, from 0 to 1 value, the higher the value the lower the AI reference to the original image (while increasing the number of iterations), personally I like low CFG high denoise heavy drawing, high CFG low denoise change the details.
If you have doubts about other parameters (e.g.
seed), please check the WebUi parameters on the model debugging page.
Under the Generate button there is an
Interrogate CLIP which when clicked will download the
CLIP for reasoning about the Prompt of the image in the current image box and filling it to the prompt.
The CLIP interrogator has two parts: one is the BLIP model, which takes on the function of decoding and reasoning about the text description from the image. The other is the CLIP model, which transforms the text into vectors (during the inference of the model).
The model used by Img2Txt is size
application see practicalguide
In img2img tab, draw a mask over a part of the image, and that part will be in-painted.
We generally have two ways of restoring images: PS and InPaint, and there are very different ways of using them.
--gradio-img2img-tool color-sketch to start up a plugin that brings in an image to be coloured (not Inpaint in this case)
PS Re-painting into Img2Img will result in a change of style, whereas Inpaint will not.
Just resize - simply resizes the source image to the target resolution, resulting in an incorrect aspect ratio
Crop and resize - resize source image preserving aspect ratio so that entirety of target resolution is occupied by it, and crop parts that stick out
Resize and fill - resize source image preserving aspect ratio so that it entirely fits target resolution, and fill empty space by rows/columns from the source image
Img2Img 3D rendering 2D
It is easier to adjust the skeleton of a 3D model than to find a sample drawing.
Pose can be combined with 3D modelling, you can use MMD related software.
If it is a real life image, you need to increase the
CFG Scale similarity and generate it together with prompts. The higher the noise reduction
denoising, the lower the relevance.
Outpainting extends the original image and repairs the blank space created. You can find this function in the img2img tab at the bottom, under Script -> Poor man's outpainting.
Outpainting, unlike normal image generation, seems to profit very much from large step count. A recipe for a good outpainting is a good prompt that matches the picture, sliders for denoising and CFG scale set to max, and step count of 50 to 100 with Euler ancestral or DPM2 ancestral samplers.
In the Inpainting tab, draw a mask over a part of the image that will be repainted.
Masked content setting determines the content placed into the masked area prior to the repair. Generally select
original to maintain potential spatial consistency, or
fill if you don't want the repaired content to inherit the original colour distribution, to use most of the image's base colour, and
latent noise to get a random colour dot pattern (to make the the generated content out of association).
|mask-sample||fill||original||latent noise||latent nothing|
The horizontal bar
mask determines the degree of blurring. original is
original, fill is
fill base and
fill takes more steps to remove the unnatural feel.
Inpaint at full resolution is a full resolution fix. By default Inpaint resizes the resulting image as a whole to the target resolution specified in the UI. When
Inpaint at full resolution is enabled, only the masked area is resized and pasted back to the original image after processing. This allows you to process large size images and allows the repair object to be rendered at a larger resolution.
There are currently several ways to carry out the redraw operation.
Draw your own mask in a web editor (
Inpaint maskedmeans redrawing the painted area,
Inpaint not maskedmeans redrawing the area outside the painting)
Erase part of the image in an external editor and upload a transparent image. The transparent area will become part of the mask.Tips: Some editors save fully transparent areas as black by default.
Change the mode (bottom right of the image) to
Upload maskand process as a separate black and white image for the mask (the white parts will be inpainted).
inpaint at full resolution appears as a black block, there may be insufficient RAM, try unload vae.
In this way, we can change the style of the character's clothing or any other detail.
Selecting the loopback script in img2img allows you to automatically feed output image as input for the next batch. Equivalent to saving output image, and replacing the input image with it. Batch count setting controls how many iterations of this you get.
Usually, when doing this, you would choose one of many images for the next iteration yourself, so the usefulness of this feature may be questionable, but I've managed to get some very nice outputs with it that I wasn't able to get otherwise.
Img2Img with low vram
As mentioned earlier, the Img2Img image quality enhancement script can be used if you encounter confusing patterns or need to produce high resolution images with low vram.
In fact I highly recommend that you use the Extras feature to replay low resolution images, it works well and is a great experience!
But if you want to use the resolution enhancement provided by the script, here's the exact procedure for Img2Img
start webui with the
Choose a smaller resolution to generate the image. Remember the resolution of the image you are generating. Once the image has been generated, copy the
Seedof the image. 3.
Once the image has been generated, check to see if you are happy with the result. If you are happy, send the image directly to img2img (click
Send to img2img)
At the bottom of the img2img screen, there is a
SD Upscaleand make the Tile overlap as small as possible.
The input box is normally filled with the original prompt when you feed the Img2img image. If you find that the prompt has changed, please fill it manually
select the appropriate
Make sure your
Heightmatch the original drawing.
Fill the Seed copied in step 2 with the Seed of img2img and generate
The Width and Height here are the sizes of the img2img images when overscaled, if they are not equal they will cause overlap problems
The SD Upscale option is in the Script section of Img2Img and is mainly used to increase the resolution.
Created: November 2, 2022 07:44:44