Kahsolt
commented
3 months ago For most regular sized images under 2k resolution, tiled VAE is all you need. For those holics wanna 4k 8k 16k, or extreme long paranomas, the latent image will still be too large to forward the UNet part, hence they'll need that tiled diffusion stuff.
PS: tiled-diff should only work properly in img2img pipeline, or txt2img with controlnet together. Otherwise, for simple txt2img, you will get those duplicated messy garbage :(
---Original--- From: @.> Date: Fri, Mar 22, 2024 01:56 AM To: @.>; Cc: @.***>; Subject: [pkuliyi2015/multidiffusion-upscaler-for-automatic1111] How tounderstand the difference between Tiled Diffusion and Tiled VAE? (Issue #367)
Apart from noise reversion and regional control, Tiled Diffusion and Tiled VAE seem to have the same effect - reducing the VRAM usage. When using text2img or img2img, I tried only using Tiled VAE and it works, whereas adding in Tiled Diffusion would create duplicate characters or other weird results. How should Tiled Diffusion be understood and used correctly? Thanks.
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>
Apart from noise reversion and regional control, Tiled Diffusion and Tiled VAE seem to have the same effect - reducing the VRAM usage. When using text2img or img2img, I tried only using Tiled VAE and it works, whereas adding in Tiled Diffusion would create duplicate characters or other weird results. How should Tiled Diffusion be understood and used correctly? Thanks.