Setting a random seed is not obvious.

[valconius]

When I set the seed's value to -1, I expect one of the following:

• the entire grid has a single randomly-generated seed
• every image in the entire grid has its own unique randomly-generated seed

I instead observe:

• the infinity grid gui shows every image's seed as -1, which is nonsense
• hunting down the text files generated alongside the images does not show seed information, which is strange (it shows it for when I set seed to a fixed value such as 12345).

I would like for the seed to be explicitly stored in the generated text file.

I would also like to have the option to make the seed either random per-grid or per-image-per-grid.

(The seed is also not in the pnginfo. I don't know the seed of my images set to generate with seed=-1.)

[mcmonkey4eva]

So as an initial reply to this issue and #14 as well:

• You're using the grid generator a bit outside the original intentions. Having randomness per-image is, well, the opposite of the intended idea of a pregenerated grid, wherein all values are consistent along a row other than one controlled variable.
• That said, it might be possible to expand the extension to support your case, it'll just take some active intentional development work to support this different usage.
• Generated seed should be in the pnginfo, if it's not that's a bug and will have to be fixed
• There isn't a text file stored alongside images. All generated text data is in one big json file and stores data per axis and per value entry within that axis - ie data unique to an individual image isn't stored except on the image itself. (And, unfortunately, for complicated unfortunate reasons, the webinterface actually doesn't have access to image metadata - refer to #6 for details on exploring that issue)
• It would be possible to dump extra files for that anyway as an optional feature if you want it I suppose

[itswhateverman]

I can see the value in these suggestions. Being able to use traditional wildcards for prompts or random seeds for the first run could speed things up, but for the comparison capability to be any value you'd want to persist it from there by writing it back to the yml or to a sidecar file so its honored on future runs.

I guess the total randomness option could have its uses in generating a large set of random images to kind of stress test a model or find seeds to iterate further on but the grid loses a lot of utility when the correlation between the rows becomes meaningless. If mass random gens is the goal an adaptation to support this use in the grid generator would probably make your tool the best suited to do so against a variety of configuration parameters.

[valconius]

I think the text file is generated by my webui settings. Therein are the generation parameters (minus the seed only if it is -1). If the seed was to be -1 then it also does not appear in pnginfo.

Presently, I put an axis that has one option for seed=constant and another for seed=-1. Before generating a grid, I manually make a few grids to find a good seed. Then, in the infinity grid, I set the random seed as a super-axis. This is a quick and easy way to check if a bad result was the fault of an embedding or hypernetwork or if that result was actually fine on another seed. Of course, rather than mindlessly generating 2x or 5x as many images by specifying additional seeds, I can just set the embed or hypernetwork as an axis and scan the random seed section to see what else popped out. It is a bit strange but think of it as fuzzing a set of parameters to vaguely correct presumptuous assumptions of success or failure. I do this when I have super-huge grids, which admittedly is only one grid so far since they take a long time to generate when I'm trying to do this.

It is also a fun setting to get infinite random funny pictures.

[mcmonkey4eva]

I fixed the bug part of this, wherein seed was -1 in pnginfo - pnginfo now has the correct seed.

[mcmonkey4eva]

I tried individual per-image metadata files in the grid tool for https://github.com/Stability-AI/StableSwarmUI and, while some of the issues mentioned before in the other thread remain present, it does work well, so I'll like port it back to here as an option.