img2img is broken

[vmobilis]

Reference image:

Command line:
./sd --model "dreamshaper_8.safetensors" -o "img2img.png" -p "green apple" --mode img2img -i "horse.jpg" --strength 0.4 --type f16

2eac844	cc92a6a	8f4ab9a

Result is almost or identical to reference.
F32 and quantized types work same.
Regression is introduced in 8f4ab9a commit.

Model version is Stable Diffusion 1.5.
https://civitai.com/models/4384?modelVersionId=128713

[stduhpf]

I can't reproduce this issue with commit 19d876e

8f4ab9a also works for me:

I don't know how this could be hapenning on your end, all I can think is that maybe it runs in inpaint mode with an empty mask? But this shouldn't be possible, if the mask is not specified, it's supposed to be initialized to fill the whole frame.

[vmobilis]

@stduhpf, I found the reason: compiler filled mask with zeroes.
If I printed few values from arr, the mask was filled with 0xFF, but otherwise the buffer seemed to be created without fill.
Depending on where I tried to copy data from arr to mask_image, I either got zeroed mask or segmentation fault.

By trials I ended up with #610.

[leejet]

There is a dangling pointer issue here, arr memory is released after going out of scope, but it is still referenced by mask_image_buffer. I fixed this issue in master, please try it if it can solve your problem.

[vmobilis]

@leejet, moving arr[] to scope of mask_image_buffer fixed the issue, thank you!

@leejet, @stduhpf, will it be acceptable to move mask creation completely to img2img part?
This will save few megabytes for txt2img generation.
Like #612.