Tiled Upscaler only works in webui but not in API

[psykokwak-com]

Hi all,
I'm trying to upscale an image using Tiled diffusion (and tiled VAE). Here are my parameters :

As you can see, it works well. The second output image comes from the upscaler.

[Tiled Diffusion] upscaling image with 4x-UltraSharp...
tiled upscale: 100%|████████████████████████████████████████████████████████████████████████████| 9/9 [00:01<00:00,  4.78it/s]

But when I apply the same parameters on the API :

    payload = {
       "init_images": [
        "myimage"
      ],
      "seed": 3073412432,
      "sampler_index": "DPM++ 3M SDE Karras",
      "steps": 6,
      "denoising_strength": 0.35,
      "cfg_scale": 1,
      "prompt": "best quality, highres, <lora:more_details:1.0> <lora:SDXLrender_v2.0:1> <lora:LCM_LoRA_Weights_SD15:1>",
      "negative_prompt": "(worst quality, low quality, normal quality:2)",
      "width": img_output_width,
      "height": img_output_height,
      "alwayson_scripts": {
        "Tiled VAE": {
          "args": [ True,       # enable
                    1024, 96,   # encoder tile size, decoder tile size
                    True,       # vae to gpu
                    True,       # fast decoder
                    True,       # fast encoder
                    False]      # color fix
        },
        "Tiled Diffusion": {
          "args": [ True,                       # enable 
                    "MultiDiffusion",           # method
                    True,                      # override size
                    True,                       # keep input size
                    img_output_width,           # image width
                    img_output_height,          # image height
                    112,                        # tile width
                    144,                        # tile height
                    4,                          # tile overlap
                    8,                          # tile batch size
                    "4x-UltraSharp", # Upscaler name
                    scale] # Scale factor
        },
      }
    }

The upscaler pass returns immediately and the output upscaled image seems to be a simple Lancoz or Nearest upscale...
I tried with different upscaler and they all returns a simple stretched resize.

Did I miss something ?

Thanks for your help

[dy-gdut]

I have encountered the same problem. Have you resolved it

[dy-gdut]

my email 1431073686@qq.com
thinks

[joel-galleri5]

facing the same issue

[aaronjolson]

Have you tried ONLY using Tiled Diffusion via the API? Does that work for you?
I am able to use Tiled Diffusion via the API using this call as my alwayson_scripts json ( I am using Python )

"alwayson_scripts": {
            "tiled diffusion": {
                "args": [
                    True,  # enable
                    "Mixture of Diffusers",  # method, eg MultiDiffusion
                    True,  # Overwrite image size
                    True,  # Keep input image size
                    1024,  # Image width
                    1024,  # Image height
                    96,  # Latent tile width
                    96,  # Latent tile height
                    48,  # Latent tile overlap
                    2,  # Latent tile batch size
                    "4x-UltraMix_Balanced",  # Upscaler
                    2,  # Scale Factor
                    True,  # Enable Noise Inversion
                    10,  # Inversion steps
                    1,  # Retouch
                    1,  # Renoise strength
                    64,  # Renoise kernel size
                    False,  # Move ControlNet tensor to CPU (if applicable)
                    False,  # Enable Control
                    False,  # Draw full canvas background
                    False,  # Causalize layers
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 1
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 2
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 3
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 4
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 5
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 6
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 7
                    False, 0, 0, 0, 0, "", "", "Background", 0, -1,  # region 8
                ]
            }
        }

It eats a ton of my VRAM though ( takes all 24gb of it and locks up my PC). I must be doing something wrong, but it also does that for me in the UI.

I have not tried using the tiled VAE with this though. I would test trying this alone first.

[joel-galleri5]

I have used tiled diffusion and control net, it works fine.

…

On Sat, 27 Jul 2024, 09:04 Aaron Olson, ***@***.***> wrote: Have you tried ONLY using Tiled Diffusion via the API? Does that work for you? I am able to use Tiled Diffusion via the API using this call as my alwayson_scripts json ( I am using Python ) "alwayson_scripts": { "tiled diffusion": { "args": [ True, # enable "Mixture of Diffusers", # method, eg MultiDiffusion True, # Overwrite image size True, # Keep input image size 1024, # Image width 1024, # Image height 96, # Latent tile width 96, # Latent tile height 48, # Latent tile overlap 2, # Latent tile batch size "4x-UltraMix_Balanced", # Upscaler 2, # Scale Factor True, # Enable Noise Inversion 10, # Inversion steps 1, # Retouch 1, # Renoise strength 64, # Renoise kernel size False, # Move ControlNet tensor to CPU (if applicable) False, # Enable Control False, # Draw full canvas background False, # Causalize layers False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 1 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 2 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 3 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 4 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 5 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 6 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 7 False, 0, 0, 0, 0, "", "", "Background", 0, -1, # region 8 ] } } It eats a ton of my VRAM though ( takes all 24gb of it and locks up my PC). I must be doing something wrong, but it also does that for me in the UI. I have not tried using the tiled VAE with this though. I would test trying this alone first. — Reply to this email directly, view it on GitHub <#365 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BJX2AWYDMCXOVZDU3MLYFBTZOMIMDAVCNFSM6AAAAABE2C46COVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENJTG4ZTCMRSHE> . You are receiving this because you commented.Message ID: <pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues/365/2253731229 @github.com>