I've been testing Ultraforge beta and what my wish in Ultraforge 1.0

First of all, I would like to mention automatic1111 in order to compare it with ultraforge so that the dev can use this to compete with automatic1111 which is a free program.
I know that there are professional experts in this forum. If there is any mistake, please help fix it. It will be useful for other users.

I have studied and used automatic1111 for a while.
It is complicated to install and requires a high learning curve.
Each time the image is rendered the quality is randomized by seed. Unable to fully control it. Even a slight edit to the prompt the image created will change.

I downloaded ultraforge beta 1.0 and tried it out.
My system has cpu i9 and gpu rtx 2070 8GB vram.
I found that the program can render images well at 512x512 pixels.
But the quality of the image in each render will be random according to the seed, similar to automatic1111.

Similarly, sometimes people's faces are distorted, extra fingers, extra limbs, and they look deformed.

As I know, SD1.5 will start rendering the first image with lower quality when using resolutions higher than the size of the images being trained in the model. For example, having additional limbs, having two heads, or having multiple people stacked on top of each other.

This means that even if you own a GPU with a lot of VRAM, you won't be able to create large images as desired. This must create an image at 512x512 with no more than 768 pixels. Most models use this resolution for training.
And the resolution in render area must not be too low so that SD1.5 can add details to the image.

When this is the case, it is necessary to render a large number of images and select good quality images to increase the quality by upscaling.

For a GPU with enough VRAM, you may check hi-res fix to increase the quality of the rendered the first image, which will get more detail and 2 times the image size.

Upscale using automatic1111 "Extras". It is a direct upscaling without adding details to the image. And will take that image and add details with img2img by inpainting the parts that want to add details by manual.

Upscale in img2img with controlnet(tile) + ultimate SD upscale is an automatic upscale. Larger and more detailed images will be obtained. Even GPUs with less VRAM can do this, but it may take more time. Sometimes you may see faint lines between the edges of the tiles.

Upscale with TileDiffusion as with ultimateSdUpscale. But there is still a problem of not being stable. Sometimes you have to restart the automatic 1111.
Each restart takes approximately 1 minute. It's really inconvenient, right?

Improving the quality of images using automatic1111 That is difficult and time consuming. Sometimes it may fail.

Right now, Ultraforge beta 1.0 can only render 512x512 images with a maximum resolution of no more than 768.
can't make batch images render.
can't upscale image with add more detail.
The advantage now is that it's easier to install. and very faster start than automatic1111.

I wish ultraforge 1.0 could upscale while adding more detail. Correct extra fingers, extra limbs, and distorted faces.
Easy to use and can use the images created for commercial.
There are other techniques such as controlling facial expressions, controlling people's poses, people interact with each other in the scene etc.

If possible, Ultraforge should have own models because most civilai.com models cannot be used for commercial use.

Quote

Kitty Ai wrote: First of all, I would like to mention automatic1111 in order to compare it with ultraforge so that the dev can use this to compete with automatic1111 which is a free program. I know that there are professional experts in this forum. If there is any mistake, please help fix it. It will be useful for other users. I have studied and used automatic1111 for a while. It is complicated to install and requires a high learning curve. Each time the image is rendered the quality is randomized by seed. Unable to fully control it. Even a slight edit to the prompt the image created will change. I downloaded ultraforge beta 1.0 and tried it out. My system has cpu i9 and gpu rtx 2070 8GB vram. I found that the program can render images well at 512x512 pixels. But the quality of the image in each render will be random according to the seed, similar to automatic1111. Similarly, sometimes people's faces are distorted, extra fingers, extra limbs, and they look deformed. As I know, SD1.5 will start rendering the first image with lower quality when using resolutions higher than the size of the images being trained in the model. For example, having additional limbs, having two heads, or having multiple people stacked on top of each other. This means that even if you own a GPU with a lot of VRAM, you won't be able to create large images as desired. This must create an image at 512x512 with no more than 768 pixels. Most models use this resolution for training. And the resolution in render area must not be too low so that SD1.5 can add details to the image. When this is the case, it is necessary to render a large number of images and select good quality images to increase the quality by upscaling. For a GPU with enough VRAM, you may check hi-res fix to increase the quality of the rendered the first image, which will get more detail and 2 times the image size. Upscale using automatic1111 "Extras". It is a direct upscaling without adding details to the image. And will take that image and add details with img2img by inpainting the parts that want to add details by manual. Upscale in img2img with controlnet(tile) + ultimate SD upscale is an automatic upscale. Larger and more detailed images will be obtained. Even GPUs with less VRAM can do this, but it may take more time. Sometimes you may see faint lines between the edges of the tiles. Upscale with TileDiffusion as with ultimateSdUpscale. But there is still a problem of not being stable. Sometimes you have to restart the automatic 1111. Each restart takes approximately 1 minute. It's really inconvenient, right? Improving the quality of images using automatic1111 That is difficult and time consuming. Sometimes it may fail. Right now, Ultraforge beta 1.0 can only render 512x512 images with a maximum resolution of no more than 768. can't make batch images render. can't upscale image with add more detail. The advantage now is that it's easier to install. and very faster start than automatic1111. I wish ultraforge 1.0 could upscale while adding more detail. Correct extra fingers, extra limbs, and distorted faces. Easy to use and can use the images created for commercial. There are other techniques such as controlling facial expressions, controlling people's poses, people interact with each other in the scene etc. Locksmith Sherman Oaks If possible, Ultraforge should have own models because most civilai.com models cannot be used for commercial use.

Ultraforge beta 1.0 simplifies installation but lacks batch rendering and detailed upscaling. Enhancements for usability and commercial usage are needed for competitive edge.