Custom seed variation

[shut pecan]

I noticed that specifying a custom seed it doesn't remain fixed across a generation. For example if I specify 900 and generate 3 images I got 900, 901, 902.

[merry halo]

well. yeah. did you want 3 exact duplicate images?

[shut pecan]

Nope. There are always subtle variations. I use ZIT, so the prompt goes through an LLM. Maybe the output from the LLM may have small differences. Even on CivitAI with a custom seed I never get exact identical images. I am even struggling to get a specific image I did: same prompt, same seed, hundreds run: never able to get EXACTLY that image. SO it is a feature anyway. Good to know. Thank you.

[merry halo]

If your inputs are identical, your outputs should be too

they will often be binary identical even

sometimes there are microdifferences that arise from gpu instability or across different machines, but these are very very small

something like a different LLM output would require a different seed to happen

seed is the magic universal "give me something slightly different on otherwise the same parameters" option

[tender hamlet]

Nope it depends on sampler, some are not deterministic

[hoary bane]

One workaround is to set the seed via <param[seed]:666> in the prompt. I think.

[shut pecan]

@tender hamlet a ray of light in the dark. Thank you. I am a noob but learned same prompt/same params/same seed = same generation. But I experienced what I was describing. I even lost my wifu... Was testing on CivitAI and magically one day the perfect style (to my taste) showed up. Man... I remixed untouched that prompt, seed, params hundred times. She never went back to me. Maybe you are pointing me: that "SAMPLER: UNDEFINED" on Civit may be the trick. I Should elaborate a grid in Swarm to run different Samplers.

[tender hamlet]

You want to avoid things like SDE, ancestral or "a" samplers or even DDIM, they tend to produce different results for each generation (same seed)

I never generated with civit so can't help u but is there not a drop feature where the parameters from your image could be read from the exif ?

As for "sampler:undefined" once again I don't know civit but it looks more like it will default to default value

[shut pecan]

The exif is that: sampler undefined. Do not know the default on Civ. On ZIT generation you can not choose the sampler. Maybe they changed it, maybe the sampler is a not deterministic one as you said.

[shut pecan]

@tender hamlet Man you are my hero! You bringed back my Wifu! Catched her with the grid testing ALL the samplers. She is there. Not perfect but 99,9%.

[tender hamlet]

Glad we were able to summon her !
Also we debunked the fake news that same parameters always give same result 😁🔥

[shut pecan]

@tender hamlet sad you are not on CivitAI, I may tip you. Great lesson learned today. I am a noob, 6 month on AI, still unable to evaluate. But today I understood the relevance of the sampler. I chose ZIT, born on Flux Schnell, tried ILL but tags hurt my mind. Never considered the importance of the sampler. On ZIT many seek variations. I am using it trying to do (lewd) comics/hentai/manhwa for fun. To me consistence instead is the key. I spend time, in company of some LLM, to build characters, detailed long prompts. Choose the seed. Now I know: a deterministic sampler. It is part of the game. On the other side I discovered AI is a nice way to create cover art for my music. There, I seek, need, crave indetermination. I feed as a prompt the song title, the key verse, the full lyrics. Looking at the outputs. Variation is key. Other than seeds now I know sampler has its role. Use cases are different. Is there a reference, a paper, without begging you to tell me, to understand which samplers are deterministic, which are not?

[tender hamlet]

Honestly u can even ask GPT
Non deterministic are older ones and I listed pretty much what you should watch out

I don't use civit for generating, only to download model and I run them locally