img2img: inpaint_only+lama throws exception (but does its job)

[Tillerz]

Hi,

like a week ago I used controlnet inpaint_only+lama to fill the area around an image to have a little space around the center object. Today I tried and it throws an error which makes no sense to me.

What I did:

• render an image in "from text"
• click "-> image"
• set a new (larger) height for the image
• enable controlnet
• select inpaint_only+lama
• select the model for it: control_v11p_sd15_inpaint [ebff9138]
• select ControlNet is more important
• select Resize and Fill
• click Generate

It generates an image, but it also throws an error before which it didn't last time:

08:44:01-222303 INFO     Loading model from cache: control_v11p_sd15_inpaint [ebff9138]
08:44:01-225804 INFO     Loading preprocessor: inpaint_only+lama
08:44:01-913313 INFO     preprocessor resolution = -1
08:44:01-934329 ERROR    Running script process:
                         X:\vlad1111\automatic\extensions-builtin\sd-webui-controlnet\scripts\controlnet.
                         py: RuntimeError
┌───────────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────────┐
│ X:\vlad1111\automatic\modules\scripts.py:406 in process                                               │
│                                                                                                                     │
│   405 │   │   │   │   args = p.per_script_args.get(script.title(), p.script_args[script.args_f                      │
│ > 406 │   │   │   │   script.process(p, *args, **kwargs)                                                            │
│   407 │   │   │   │   s.append(f'{script.title()}:{round(time.time()-t0, 2)}s')                                     │
│                                                                                                                     │
│ X:\vlad1111\automatic\extensions-builtin\sd-webui-controlnet\scripts\controlnet.py:831 in process     │
│                                                                                                                     │
│    830 │   │   │   logger.debug(f"Use numpy seed {seed}.")                                                          │
│ >  831 │   │   │   detected_map, is_image = preprocessor(                                                           │
│    832 │   │   │   │   input_image,                                                                                 │
│                                                                                                                     │
│                                               ... 3 frames hidden ...                                               │
│                                                                                                                     │
│ X:\vlad1111\automatic\extensions-builtin\sd-webui-controlnet\scripts\processor.py:489 in lama_inpaint │
│                                                                                                                     │
│   488 │   # applied auto inversion                                                                                  │
│ > 489 │   prd_color = model_lama(img_res)                                                                           │
│   490 │   prd_color = remove_pad(prd_color)                                                                         │
│                                                                                                                     │
│ X:\vlad1111\automatic\extensions-builtin\sd-webui-controlnet\annotator\lama\__init__.py:51 in         │
│ __call__                                                                                                            │
│                                                                                                                     │
│   50 │   │   │   mask = (mask > 0.5).float()                                                                        │
│ > 51 │   │   │   color = color * (1 - mask)                                                                         │
│   52 │   │   │   image_feed = torch.cat([color, mask], dim=2)                                                       │
└─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
RuntimeError: The size of tensor a (3) must match the size of tensor b (0) at non-singleton dimension 2
08:44:02-593506 INFO     Loaded SwinIR model: X:\vlad1111\automatic\models\SwinIR\SwinIR_4x.pth
                         param=params_ema
Upscaling SwinIR 100% ----------------------------------------------------------- 36/36  [ 0:00:12 < 0:00:00 , 3 it/s ]
Progress 5.19it/s ---------------------------------------- 100% 0:00:00 0:00:07```

Python 3.10.9 on Windows
Version: 875d0db Sun Jul 2 08:26:51 2023 -0400
nVidia CUDA toolkit detected
Torch 2.0.1+cu118
Torch backend: nVidia CUDA 11.8 cuDNN 8700

[Tillerz]

Ok, correction to the above:

1. It does NOT its job, when it throws the error. It cannot load stuff (see 2.) and then just continues with DEFAULT inpaint or whatever, ignoring the inpaint+lama selection.

2. Apparently, inpaint+lama only works, if you upload/drop the image into the controlnet image box, too. This fixes the error. So this is not really a bug, just a usability annoyance. :D

Would be handy if you could tell the UI to use the image as ControlNet input, too.