TeaCrab
9c3af9fc8a
* RandomLora & RandomModel
* Walks the `models/checkpoints` and `models/loras` folder for `.safetensors` files and uses Regex patterns to filter them down, then randomly select one to apply.
* If the `pattern` is changed, filtering and randomization will be triggered next time the node is executed.
* `every` parameter allows the selected model to loop for a certain number of times before next randomization.
* `pause` holds off randomization and keep using the same model for generation indefinitely.
* `skip` will randomly select another model ignoring the `every` parameter but the looping will continue to count up.
* `RESULT` output will show in text format the current number of loops, which model is currently selected, which model has ran and how many times, and lists all the models that hasn't ran. Use `Preview as Text` or `Show Text` node to access.
* Models are pooled - loaded to RAM from HDD/SSDs, this causes RAM to be filled to the brim, however, Python seem to be able to automatically release memory for new ones
* If previously loaded models are selected again, it is already in the RAM, so no need to access SSD or HDD for it again.
* This saves the tiny bit of load time when using SSD, but could potentially save a lot of loading time for HDD.
* Eventually this could be used for cycling through multiple designated models to compare their differences.
ComfyUI-TeaNodes
Adds a few new nodes:
Image Equalization CLAHE (This node can't be any more Kornia)
- When images have their histogram smoothly distributed, I'd say it gives ControlNet preprocessor an easier time.
- Works great when image BG is removed.
Image Size Approximation
- Works based on pixel count that retains Image ratio.
- This algorithm is really dumbly written, but it works.
- Great for stabilizing the speed of image-to-image generation.
Image Resize
- Takes size tuple from Size Approximation Node.
- Seriously, image size always has 2 numbers, why can't it fit through a single wire?
- It also defaults to a node socket instead of having to convert it into input from a widget.
Image Scale
- Simply multiplies image size by a factor.
- 20240425: Now scales smoothly when multiplied by non-power-of-2 factors - use LANCZOS option now
- Have an easier time saving in-process scrap images at half or quarter resolution.
Crop To
- Crop an image to the same size of the reference image
- Hate some of those nodes that complains about bad tensor dimensions for whatever reasons? Fear no longer.
KorniaGamma
- This works like an easy/complex/idk brightness/contrast/level adjustment node
- Image processing power must be utilized for better results.
Random & Non-Random Color Fill
- Fill based on reference image size
- Randomness defined by original color and a variance in hue, saturation and value
- Hopefully the color input field works for most color string formats.