Is there a way to increase the step size of the seed when generating in batch?

[ash0080]

The default step size for the seed is currently 1, but similar seeds tend to produce similar outputs. If I want the generated images to be as different as possible within a batch, then a step size of 1 is not very convenient.
So my question is whether it's possible to modify the step size, and even randomly vary this step size within a certain range?

[ThereforeGames]

Hi @ash0080,

You can do this with the Unprompted extension.

Batch seeds are stored into the array variable all_seeds. Use a [for] loop to update the values as needed.

[sets step=100 _new] [# Change step size as desired, or use a random block]
[if batch_index=0] [# We only need to process the seeds array once]
	[for i=1 "i < {length all_seeds}" "i+1"]
		[set new_seed][eval][get seed]+[get step]*[get i][/eval][/set]
		[array all_seeds i="{get new_seed}"]
		[log]Setting seed of image [get i] to [get new_seed][/log]
	[/for]
[/if]

For your convenience, I went ahead and added this as a template in the latest version of Unprompted. To use it, simply insert [file common/examples/batch_seeds_custom_step_size] anywhere in your prompt.

You can alter the step size like this:

[file common/examples/batch_seeds_custom_step_size step=55555]

Or randomize it like this:

[file common/examples/batch_seeds_custom_step_size step="{random _min=100 _max=500}"]

Good luck!

[ash0080]

Wow, you are so kind. I will try it now. Thank you!

[w-e-w]

but similar seeds tend to produce similar outputs

most likely you're just imagining things

[w-e-w]

random number generators are are specifically designed so that there is no relationship between one seed and the next (within it's designed use case)

[w-e-w]

if you're actually able to see relationship between one seed and another then you have a topic for a paper
and this paper will have a great impact on RNG

[ash0080]

Before you rush to present any results, I think you should at least understand the application scenario. Your example is simply a different application scenario. Can randomization on a base model be the same as randomization on a small LoRa dataset? Maybe, but would someone be able to notice the difference when randomizing on such a large base dataset? It's a bit funny.

If you want to test it, I can tell you that you can draw a triangle, a square, a circle, and a pentagram, combine them to form a LoRa, and then randomly select them using only the "shape" keyword. You will see the pattern.

At least I did this experiment instead of talking nonsense based on experience. Who isn't a programmer here?
Also, if the variation of seed step size does not affect the results, please explain why variation seed can control the step size?
You mentioned the paper. What I want to say is that at least so far, I haven't seen any paper that explains that the seeds in SD is randomly distributed in a normal distribution. So, no one can give a definitive conclusion, can you?

Also, I'm just asking for a feature request, and if the feature request itself doesn't make sense, I think it's fine to discuss it, but it's rude to start mocking me as a matter of course when you don't know anything about what I'm talking.

最后，请收起你简中的那套说话习惯，这里又不是微博知乎，而我至少现在还是civitai排名第一的写作者

[w-e-w]

好 現在回溯下

首先你發問

The default step size for the seed is currently 1, but similar seeds tend to produce similar outputs. If I want the generated images to be as different as possible within a batch, then a step size of 1 is not very convenient.
So my question is whether it's possible to modify the step size, and even randomly vary this step size within a certain range?

我的回覆

首先 你說出了個你觀察到的現象 但並沒提供任何依據 參照 驗證方式
你提出了個現象與多人驗證的解果核經驗不符合

所以這裡判斷 你絕對是產生錯覺 所以回復

most likely you're just imagining things
我認為這只是陳述我的看法 而且確實也有別產生這樣的錯覺的先例

你的回覆

I think it's better to try it yourself first when you mock others (including the friend who downvoted).

我是嘗試後才做最初的回覆的，你沒提供你的實驗法所以我只能根據"預設 做基本的測試 結論 無關聯性"

when you mock others

我認為你是解讀太多 產生我侮辱你的錯覺

(including the friend who downvoted).

又不是我點👎的，關我啥事?

I have been working on a series of LoRa experiments that rely on batch. I have done many comparative experiments. When I use batch, such as batch=9, the randomness distribution on the original dataset is not so ideal. However, if I use count=9, the situation is significantly different. Currently, I have solved this problem by adding a random meaningless string through dynamic prompts.
But I think directly manipulating the seed is a better idea.

很好，你有做實驗，問題是還是沒展示出你實驗結果，所以無法驗證，步驟也模糊 所以我這也無法驗證

I don't think this is a field that everyone is familiar with, and I haven't seen anyone else making a LoRa like I did. So, maybe your experience has limited your imagination? 😮‍💨

確實 這部分是很新的技術 很多東西還沒被完全了解，確實有可能經驗框住

Looking at it from a different angle, even if we don't talk about the randomness of the seeds, I think it's reasonable to request an increase in batch step size by a non-incremental step, because this is an existing feature in ComfyUI. 🙄

問題是在我們這方看來，根本沒有此功能的必要，要加入這功能很簡單，但除非能被確認是"重要功能"不然很難說服要把額外的功能加入這個已經很難維護的複雜系統中，而且是在如果真有需要也已經有插建可實現的前提下

https://civitai.com/models/116525/game-dev-tools-01-grid

你貼了個模型連結 沒說想要表什麼

我的回覆

我給出了兩組16張圖 分別是用不同模型在連號seed和隨機的 根據觀察 無法區分出哪組是連號 哪組隨機 所以結論 沒變
並上傳了圖片

"most likely you're just imagining things"
believe it or not this is not a criticism this is a statement of fact

we humans have a tendency of seeing patterns where there aren't

like seeing shapes of faces or animals in clouds

we humans are very very good at this seeing patterns to point that we can see patterns in just about anything

technically in this case it since pytorch RNG is pseudo random "there is a patterns", but the patterns should be practically impossible to detect in our use case

random number generators are are specifically designed so that there is no relationship between one seed and the next (within it's designed use case)

我把以前電腦課上學的基礎總結貌似也被當作是在嘲諷

if you're actually able to see relationship between one seed and another then you have a topic for a paper
and this paper will have a great impact on RNG

這句算是把我一個教授的套用過來，如果你發掘了個問題 特別還是這種重大問題 這是可以發論文的

你剛才的回覆

Before you rush to present any results, I think you should at least understand the application scenario. Your example is simply a different application scenario. Can randomization on a base model be the same as randomization on a small LoRa dataset? Maybe, but would someone be able to notice the difference when randomizing on such a large base dataset? It's a bit funny.

那你也得把你的實驗放發說請出，你上面給的說明非常模糊，無法構成一套流程 所以當然我只能用我自己個發法驗證

If you want to test it, I can tell you that you can draw a triangle, a square, a circle, and a pentagram, combine them to form a LoRa, and then randomly select them using only the "shape" keyword. You will see the pattern.

這又是在說什麼?
你是說要自行繪製多個三角方形圓形五角形的圖片 然後練個lora嗎? prompt 統一用 shape 不用各自的型形狀名稱?
還是prompt用triangle shape square shape circle shape pentagon shape 去訓練
然後最後用webui跑圖只用prompt 只用shape "隨機"看看會出來什麼?
總之我是看不懂你要我怎樣再現你所觀察到的現象

At least I did this experiment instead of talking nonsense based on experience.

還是請停工詳細再現步驟 最好連 模型的 所用的配指等等全都的說清楚 不然要人怎樣再現

Who isn't a programmer here?

你大概是不請處有多少非程序原在這個repo中留言 很多甚至是新創的沒幾天的號
看別人留言沒事又不會去點進profile做調查知道你是誰

在這repo跟很多人說某個蓋念是需要從基礎說的 太多非專業人士 如果有人不知道你是誰 然後跟你用講基礎 你覺得你被侮辱到了 我覺得這是你的問題

Also, if the variation of seed step size does not affect the results, please explain why variation seed can control the step size?

variation of seed step variation seed 說請處你在說的是什麼功能 如果是再說https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#variations 那這就跟seed 本身沒啥關西

You mentioned the paper. What I want to say is that at least so far, I haven't seen any paper that explains that the seeds in SD is randomly distributed in a normal distribution. So, no one can give a definitive conclusion, can you?

大概又是溝通誤解 我是在說 如果你真的能在目前用的RNG算法透過宏觀觀察找出關聯性 那這可以是個論文(paper)的主題

Also, I'm just asking for a feature request, and if the feature request itself doesn't make sense,

對沒錯 你是在提出功能的請求
所以如果你真希望此功能被加入 你需要做的事讓人理解到為何需要此功能

I think it's fine to discuss it, but it's rude to start mocking me as a matter of course when you don't know anything about what I'm talking.

我覺得你想太多 沒事幹嘛侮辱你

最后，请收起你简中的那套说话习惯，这里不是微博，

啥 哪套? 簡中? 簡體中文?? 我在這之前是用英文你溝通

我至少现在还是civitai排名第一的写作者

要發現這邊跟上面不知為何張貼的civitai網頁的作者 竟然是跟我在對話的這人使用了同一串字符還真是不容易

請你回答

1. 詳細 實驗步驟 包含你所指用的模型 設置 過程 詳細到一個完全沒概念的人可以造作
2. 提供你自己觀察到的現象 包含圖片
3. 你是哪點看到我在侮辱你

[ash0080]

Please, you're just a collaborator, and someone else has already answered my question. I just didn't come back in time to close it. Do I need to answer your questions? As a collaborator, your first sentence was sarcastic, and now you finally remember to ask me to submit a case study? You enjoy it, but I don't have the time or obligation to answer your questions. Your stress reaction makes me feel like I don't really have the desire to communicate with you. Thank you.