UltraFlux, a high quality VAE decoder for Flux.1

by
0 comments
UltraFlux, a high quality VAE decoder for Flux.1

This is a newly trained VAE for Flux.1 that has been trained on high resolution 4k images. This means that when you generate large 4k resolution images using Flux.1 this model VAE decoder will produce higher quality results than the original VAE provided by Black Forest Labs.

Its description is provided by the manufacturer “ultraflux is the flux-based diffusion transformer for native 4k With text-to-image generation Consistent quality across different aspect ratios (Square, portrait, ultra-wide, etc.).
It combines a 4K-optimized DiT backbone with curated MultiAspect-4K-1M Dataset, VAE post-training, and 4K-aware objectives Data, architecture and loss are co-designed rather than managed separately.,

Let me just say that this is definitely an improvement over the images produced by Flux.1. This is without using any LoRA or descriptors to begin with. When the flux generated latent is decoded using this VAE model, the results are clear and contain a lot of fine details.

You can download VAE from them embrace home page

Place the download file in the ComfyUImodelsvae folder, make it something like UltraFlux-VAE.safetensors so you know it’s a separate file.

The VAE simply decodes the same latent image that is created by the sampler and decodes it using this custom VAE which produces a higher quality image that is sharper and has more details than others. The workflow used in ComfyUI is the original workflow that was released for flux.1 dev and I have simply added an additional decoder node, loaded a custom vae and added the preview image.

To compare I use the Image Compare node from RGThree and to animate the result I use my WWAA Before After Animator which is combined using the Video Combine node to save as a video.

This is a workflow preview.

Sample results as shown in the video above.

Image details are clear and there is no typical flux softness at all. You just have a clean and clear image to start with, now if you want to take it further just add some detailed LoRA for Flux.1 and you are on to a winner!

If you would like to support our site please consider buying us ko-fi, take a product Or subscribeNeed a fast GPU, get access to the fastest GPU for less than $1 per hour RunPod.io

Related Articles

Leave a Comment