This is a newly trained VAE for Subflux 1 trained on high resolution 4K images. This means that when you produce large 4K resolution images using the Flux 1, this Model V decoder will produce a high-quality result provided by the local VAE Black Forest Labs.
Its details are as provided by the creator”Ultraflux is a flux-based diffusion transformer Native 4K Text-to-image generation with Consistent quality across varying aspect ratios (Square, Portrait, Ultra Wide, etc.).
It combines with a 4K-optimized DIT backbone Multi-Aspect -4K-1M dataset, VAE post-training, and 4K-familiar objectives so that Figures, architecture and damage are designed interactively rather than tuned in isolation.“
All I have to say is that it is definitely an improvement over the images produced by the Flux 1 right off the bat. This is without using any LORA or specifications to begin with. The results are crisp and contain a lot of detail when the flow-induced latency is decoded using this VAE model.
You can download VAE from them Hug Face Page
Place the download file in the Comfyui\Models\VAE folder, give it a MAC like ultraflux-Vae.safetensers so you know it’s a different file.
VAE simply decodes the same lingering image as produced by the sampler and decodes it using this custom VAE that produces a higher quality image that is sharper and contains more details than others. The workflow used in Comfy is the original workflow that was released for Flux 1.
To do the comparison I use the Image Compare node with RGTHERE and for the resulting animation I use my WWAA after Animator which is then combined using the Video Combine node to save the video.
Sample results as shown in the above video.
Image details are clear and the general flow is not soft at all. All you have is a clean and crisp image to start with, now if you want to take it further just add some detail Laura for Flux 1 and you’re on to a winner!
If you would like to support our site, please consider buying us Co-Fifor , for , for , . Grab a product or Subscribe. Need a faster GPU, get access to the fastest GPUs for less than $1 per hour runpod.io