■onetrainer_config:Training Presets

■You’ll also need the following file for training—it’s easy to overlook. Please specify it in the Training tab:

●effnet_encoder.safetensors

If you want to train with Cascade 1B or a single-file model, set that single file in the prior model field within the Training tab.
Also, models are available in float32 and bf16, but since the model itself was trained in bf16, there shouldn’t be much difference whichever you choose.

●stage_c.safetensors(3.6B)
●stage_c_lite.safetensors(1B)

All the data are available at the URL below.
https://huggingface.co/stabilityai/stable-cascade

■cascade_comfyui_workflow.json:
I’m using ClownsharkBatwing’s workflow—one of the pioneers of Cascade—almost as is.
The workflow includes clear, detailed instructions for downloading the models needed for inference.
Please install the following required extensions.The inference resolution is set to match the base model. 
If you fine-tune at 1024 px (e.g., higher training resolution), you can run inference at even higher resolutions.

●RES4LYF:
https://github.com/ClownsharkBatwing/RES4LYF

●UltraCascade:
https://github.com/ClownsharkBatwing/UltraCascade

●Uncond-Zero-for-ComfyUI:This is Auto CFG. It isn’t strictly necessary, so you don’t need to install it.
https://github.com/Extraltodeus/Uncond-Zero-for-ComfyUI

■modules>WuerstchenBaseDataLoader.py:
This file contains my modified values for Cascade’s internal image downscaling.
If you feel this change is useful, either replace your file with this one or edit the internal values yourself.

In OneTrainer, when training Cascade, images are internally downscaled to 0.75×.
So if you set 1024 px in the UI, the actual training runs at 768 px.
This is a UI-oriented choice to avoid confusion, letting users think “train at 1024 px, infer at 1024 px,” which feels more intuitive.

However, if you truly want to train at the full 1024 px, you have to set a higher resolution in the UI, which is inconvenient.
Therefore, change the following line from factor=0.75 to factor=1.0 to disable downscaling.

■"downscale_image = ScaleImage(in_name='image', out_name='image', factor=0.75)"

With this change, training will use the actual, specified resolution.

If you’re unsure, change the value and then rebuild the cache—if the cache size increases, it worked. After success, setting 768 px should produce a cache size that’s almost the same as the pre-change cache size at 1024 px. 
This confirms that 1024 px was in fact being downscaled to 0.75×.

"downscale_image = ScaleImage(in_name='image', out_name='image', factor=0.75)"