Comfy UI, simple Txt To Image setup

 

The default Comfy UI graph looks like this.

Load Checkpoint: Here's where you specify the model, eg Stable Diffusion 1.5, XL, ToonYou whatever

Clip Text Encode Prompt: your text prompts (the red and green are the same type of node, just coloured), both of them are connected to the Load Checkpoint's Clip output

Empty Latent Image: Here you specify the resolution of the image & batch size, that we'll be working with

These 3 all feed into the KSampler, as shown above. The KSampler lets you specify Seed number, sampling steps, type of sampler (Euler, etc), how much denoise to use.

KSampler & the VAE output of the initial Load Checkpoint feed into the VAE Decode node. Finally the Image output from this gets fed into a Save Image.

VAE - it is worth noting that most checkpoints based on SD 1.5 has a VAE built in, so you can feed that straight to the VAE Decoder, from the Load Checkpoint. I had rather drab desaturated results when I used a separate VAE that was recommended vs the more vibrant expected results from the built in one. So this last sentence is not entirely true. I think it might be AnimateDiff's loop feature causing this!



Comments

Post a Comment

Popular posts from this blog

Comfy UI AnimateDiff BASIC setup

 Here's a screenshot of a basic ComfyUI AnimateDiff setup/workflow. The setup is largely the same as the previous post covering the default Txt2Img one. The new nodes to pay attention to are the Animate Diff Loader and the Uniform Context Options. Also note that the Empty Latent image has been set to 32 in the batch size . This is the value that will dictate the length of your animation - NOT the number set in Uniform Context Options! Uniform Context Options is what lets us chain together "unlimited" AnimateDiffs. With standard setup, you're limited to 36 frames (at time of writing) but here we can set the length of each "block" of AnimateDiff (I leave it at 16), the overlap of each block and whether we want to have the animation a closed loop or not. I find that closing the loop produces washed out results, I'll post an update if I find a solution/workaround. Context stride by default is set to 1 - this is OFF. I haven't found decent explanations ...
Read more

face swap "deep fake" with stable diffusion and roop, in Automatic1111

 https://youtu.be/AAUSiihcpj0  install Roop in extensions, restart UI. For stills - enable Roop, load in target face. Prompt SD as usual (eg. dutch oil painting  of man staring ) For sequence - switch to img2img tab - batch mode  - enter source directory. Specify correct aspect ratio of image /pixel size. Reduce denoise strength to zero, as we don't want the model to do anything. Enable Roop, load target face.. hit process. Quite simple!
Read more