alecjacobson / computer-graphics-text-to-image Goto Github PK

The queue and time to generate an image is taking over 20 minutes - is there anything that can be done to make the process faster?

For bias.zip, do we generate 4 sets of images and identify a bias in each set or do we generate one set of images and identify 4 biases from that set of images?

I generated an image a couple days ago and when I came back to working today (for open-ended), I ran the same prompt in the same seed and it gave me a different image which kind of messes up the progression I had going. Is this normal? When I regenerated on the same seed with the same prompt before the same image was generated.. Is there a way to get the old generation I had before?

This is kind of wired that when I tried to run with same seed and same prompt, I am getting totally different result from yesterday, is this normal or something goes wrong?

does the movie poster need to have a title? Its not part of the constraints

For movie posters:

1. Are we supposed to set the title, genre, and constraints based on a movie we choose or are supposed to choose from the movies in the README?

2. Is the goal to create prompts that output movie posters that fit the title, genre, and constraints?

My movie posters are not able to satisfy all the constraints I give in the prompt. It seems like at least one constraint is not satisfied in the image everytime, although I have expressed it in the prompt in a specific and detailed manner.

For example, if I give the prompt "aquaman with underwater city behind". It will give an image with aquaman and the underwater city. But when I add more to the prompt "aquaman with long hair holding a golden trident with an underwater city behind". Its just disregards the underwater city, and theres no way I can get it to show up again. Am i doing something wrong?

Also, a lot of times, aquaman's face or part of his body is cropped out of frame, eventhough I have specified "full body" in the prompt, not sure what I can do to fix that.

Any tips would help.

As far as I know, the best way to generate a detailed image is by generating a great amount of images at once and then selecting the best one among them. However, we are only allowed to generate 4 images at a time on the school computer. It would take a great amount of time just clicking on the screen once in a while. Can we use stable diffusion web UI on our own computer with the same model to generate the images we need for the assignment? I have some experience of doing that on my own and I know how to do it. If I'm to run it on my own computer, what should I notice so that my assignment can be graded properly (i.e. can we use image-to-image generation, change some settings, or even use Loras?)

Task 4 want 6 image about a (un-popular) character in same style
and after the 30 min+ generating, when we want another image with the same prompt, the model is already changed, and we get extremely different character


this is 2 image I got with same seed, almost same prompt

Hi, I just saw the update that Assignment 9 will be extended to Friday. Thanks for that. However I would just like to confirm that, if we can still apply the rest of our late days to Assignment 9. Lets say I have 3 late days left, and can I use it to extend the assignment to next Monday? Thanks in advance.

I was trying to submit my first task bias.zip (having passed the validation) but Markus gives me this error message:

(it is 14.4MB for reference)

What can I do about this, especially for later tasks that will have more images in its .zip file?

I've been trying a bunch of different prompts for the movie poster task and I have not been even close to producing something "visually attractive" that fulfills the constraints.

Any tips for how these prompts should look like or what I should avoid? I've tried giving detailed descriptions of the scene and also tried listing descriptors with no luck so far.

how many images should we generate for each bias? Is 4 images enough per bias?

Are we supposed find biases that are present in the image generator, but not in the real world? Or can we also identify biases present in both the image generator and the real world?

Do we need to cite where we got our statistics from?

I've been trying to do this assignment for a while, but the images generation takes a very long time, even though I'm the only one in the queue (I can't imagine how worse it'll get as the deadline approaches), and I keep getting errors.

Any advice would be appreciated!