Wanted to see what effect training Stable Cascade Stage B (a.k.a. the decoder) would have. The effect is very subtle, but it seems that this model is responsible for the really fine details in the final image, and somehow my very naive training seems to have improved the model noticeably. My trained version produces images that appear sharper, less blurry. The effect is especially noticeable in images that have fine black lines like sketch-type drawings.The dataset I used is about 5k images, lots of different styles, including some NSFW, but I don't think it really matters much in Stage B since it isn't really influential in the overall content of the image or the style; in fact, Stage B would probably benefit most from being trained exclusively on highly detailed photos and zero illustrations (except maybe if you wanted a pipeline totally dedicated to a specific style). I will continue to experiment.I trained these using the official Stability AI source code, with some modifications to make it work on my GPUs.Getting this model to load is an exercise left up to the downloader. I got it working using diffusers StableCascadeUnet.from_single_file (this is needed as it contains the conversion logic). From there, you can create a pipeline and run inference or save the pipeline to disk in HF format. If you want to load this in ComfyUI, I think HF format may work, or there might be a conversion script somewhere, I don't know, I haven't used Comfy.Note that in the showcase images, the first image was generated using this model, and the second image was generated with the official Stable Cascade Stage B. Notice the details of the individual hairs are visible and sharp in my version, while the original version is a little more blurry.

Stable Cascade Stage B Slightly Improved

First attempt. This is highly experimental, but the results seem noticeably better.Sample Images were produced using AltCascade as Stage C (the prior), and this as Stage B (the decoder). All other components were original Stable Cascade.

Files with SHA256: 4e99adaf00791a3331104825c74ee2f6252332af3c58ed09856fac3ff3dae31b