I do architectural renderings and often need to generate large images. That soft, mushy blur when Flux.1’s native VAE decodes at 4K has always been annoying—details just get flattened out. Recently I tried a VAE called UltraFlux, which is a decoder retrained on 4K high-res material, specifically aimed at large resolutions. Super easy to use: just download the file from its Hugging Face page, drop it into ComfyUI\models\vae. I renamed it UltraFlux-VAE.safetensors to keep things straight. No major workflow changes—just the usual Flux.1 Dev setup, but after the sampler, add a VAE Decode node, hook up the custom VAE, then preview.
Same latent, just swapped the decoder, and the output is noticeably sharper. Reflections on glass curtain walls, brick textures, those edge details actually hold up—Flux’s signature softness is almost gone. I haven’t even added any detailer LoRA yet, and bare output at this level is already good enough. If you want to push it further, just slap on a Flux detail LoRA after. For comparison, I used rgthree’s Image Compare node—sliding the slider back and forth makes the difference obvious.