• Gears

    Render best image with Stable Diffusion

    I am running Automatic 1111 on Windows, of course I want to render the best images with Stable Diffusion like everybody else. But as a matter of fact, there is no ‘good’ or ‘bad’, as far as we talk about the result.

    There is not such a thing as the ‘best’ image at all, it is all about what my target is. Or say, what do I want to get from Stable Diffusion. The answer is decent details and sharpness with reasonable resolution.

    Another key factor needs to be considered is hardware. My AMD Radeon RT 6700 XT is not that powerful which could act like James Bond or Jason Bourne to save my world, it can only work in the Liam Neeson way otherwise it won’t deliver ideal result or may even end up in disaster.

    When I try 512×512 or 768×768 pictures, it works perfectly by keeping both speed and result in a balanced way. But to render 1024×1024 would die hard, due to the restirction of GPU. Anyhow I do need higher resolution pictures, and at the same time with more details and good sharpness.

    Stable Diffusion is in fact the science of practice, different method or process may resullt in quite different output. The key could be a proper work flow. Reasonable process may enable me to achieve my target.

    I am still working on SD1.5 family models. I played around a bit and found out below work flow fits my demand.

    I started with 512×512 resolution, and used Hires.fix to upscale the picture by 1.5x, hence I got result in 768×768 then could continue from there.

    512×512 upscaled by Hires. fix to 768×768

    The 768×768 picture is overall good, but I am expecting more details. Detail Tweeker LoRA willl do the job. By the way, there are 2 similar LoRAs, this one is for SD1.5 models, there is another one ‘Detail Tweeker XL’, with ‘XL’ in name, is for SDXL family models.

    Since I only need to add details to the original picture, I do not want to make this step too complicated. I only employed Detaill Tweeker LoRA and I am happy with the rendered image.

    add details with Detail Tweeker LoRA

    Now I want more sharpness, and one thing at a time. I used 4x Ultra Upscaler to bring sharpness to the existing 768×768 image. Since I only want to bring more sharpnes, hence upscale ratio can be set to 1 to prevent the image from being upscaled.

    after applying 4x Ultra Upscaler

    Last thing is to upscale the picture. Since it is only a test, I tried to upscale to 1152×1152. Below is the outcome.

    upscaled to 1152×1152

  • Opinion

    Is Payment System in Germany Secured?

    In addition to vast numbers of online shops, more and more retail shops in different categories, restaurants and cafes accept electronic payment other than payment by debit card or credit card. One can conveniently choose to pay by credit card, debit card, even to utilize tools from payment service providers, such as PayPal, Apple Pay or Google Pay.

    In Germany, it doesn’t matter which institution issued your debit or credit card, it is always either Master Card or Visa sitting behind it. As for payment providers, you can hardly find one beyond PayPal, Apple Pay or Google Pay. To make it simple, the market is dominated by giants from the United States.

    And don’t forget SWIFT. In case there is turbulence in market stability and the USA payment service providers fail to deliver their products to the German market, it is hard to imagine how consumers in Germany could handle the turmoil. Without a German local service provider maintaining a decent market share, is payment system in Germany secured? Maybe for now, yes.

  • Gears

    Stable Diffusion on AMD GPU

    I installed Stable Diffusion Web UI on my PC, which has AMD Radeon RT 6700 XT graphic card. When testing, I found out that UI is not utilzing GPU, but only CPU.

    After trying different settings, CPU utilizatzion ratio dropped down to 75-78%, Stable Diffusion is still not using GPU.

    Now below method solved my problem, tried to render pictures with Automatic1111, GPU utilization rate is 99% when rendering, whilst CPU is below 25%.

    I am happy to share the process which works pretty well on my PC, hope this could be helpful to those who encounter the same problem as mine.

    Suppose AMD graphic card driver is already up to date, Pathon 3.10.x (add to PATH) and Git Bash are installed.

    Create a new folder under root, like c:\sd, do NOT put it under folders like Desktop, My Documents or so.

    Install web UI under c:\sd folder: “git clone https://github.com/lshqqytiger/stable-diffusion-webui-directml && cd stable-diffusion-webui-directml && git submodule init && git submodule update”.

    Edit “webui-user.bat” file with editor, insert “COMMANDLINE_ARGS=–skip-torch-cuda-test –opt-sub-quad-attention –disable-nan-check –opt-sdp-attention –precision full –no-half –use-zluda –skip-ort” and save the file.

    Run webui-user.bat, if error occurs, this is normal at the current stage. Just forget it and close web UI and Terminal.

    Download and install AMD HIP SDK 6.2.4.

    Download the latest Zluda files for ROCm.

    Create c:\zluda folder, extract all file from zluda-windows-rocm6-amd64.zip, paste everything under c:\zluda.

    Open c:\Program Files\AMD\ROCm\6.2\bin\rocblas folder and find “library” sub-folder, rename “library” to something like “library-backup” or similar, so that one won’t forget it. In case later on we may need the files again.

    Download library for hip sdk based on gfx version, RT 6700 XT = GFX1031. Extract files then put extracted file “rocblas.dll” and folder “library” under c:\Program Files\AMD\ROCm\6.2\bin\rocblas as they are, now we have a new “library” folder which contains downloaded lib files we need.

    Add c:\zluda folder and %HIP_PATH%bin to PATH, restart PC.

    Now we run webui-user.bat, to render image for the first time could or could not take longer, but as from the second image, it is charmful.

Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.