Skip to main content ↑ ↓
Topic: Welcome to AI Image Central! (Read 11870 times) previous topic - next topic

Re: Welcome to AI Image Central!

Reply #30 –
New update:

- Added sampler and scheduler options to Flux models.

Here is a little guide for Flux model selection I found on Reddit:

These are the only scheduler/sampler combinations worth the time with Flux-dev-fp8.

🟢=Good 🟡= Almost good 🔴= Really bad!

Here I have compared all sampler/scheduler combinations by speed for flux-dev-fp8 and it's apparent that scheduler doesn't
change this much, but the sampler does. The fastest ones are DPM++ 2M and Euler and the slowest one is HeunPP2.

From the following analysis it's clear that the scheduler Beta consistently delivers the best images of the samplers.
The runner-up will be the Normal scheduler!

SGM Uniform: This sampler consistently produced clear, well-lit images with balanced sharpness. However, the overall mood and
cinematic quality were often lacking compared to other samplers. It’s great for crispness and technical accuracy but doesn't add
much dramatic flair.

Simple: The Simple sampler performed adequately but didn't excel in either sharpness or atmosphere. The images had good
balance, but the results were often less vibrant or dynamic. It’s a solid, consistent performer without any extremes in quality or
mood.

Normal: The Normal sampler frequently produced vibrant, sharp images with good lighting and atmosphere. It was one of the
stronger performers, especially in creating dynamic lighting, particularly in portraits and scenes involving cars.
It’s a solid choice for a balance of mood and clarity.

DDIM: DDIM was strong in atmospheric and cinematic results, but it often came at the cost of sharpness. The mood it created,
especially in scenes with fog or dramatic lighting, was a strong point. However, if you prioritize sharpness and fine detail, DDIM
occasionally fell short.

Beta: Beta consistently delivered the best overall results. The lighting was dynamic, the mood was cinematic, and the details
remained sharp. Whether it was the portrait, the orange, the fisherman, or the SUV scenes, Beta created images that were both
technically strong and atmospherically rich. It’s clearly the top performer across the board.

When it comes to which sampler is the best it's not as easy. Mostly because it's in the eye of the beholder. I believe this should be
guidance enough to know what to try.

Re: Welcome to AI Image Central!

Reply #31 –
Found a chart for SD 3.5 that shows good combinations for Stable Diffusion 3.5 samplers and schedulers:

Re: Welcome to AI Image Central!

Reply #32 –
New update:

- Added sampler and scheduler options to SDXL models.

Also take note, the samplers and schedulers I believe work best are near the top of the list but I have included many that mostly work for the different models.
A couple combinations will give bad results.  Check the charts in the previous messages.

Re: Welcome to AI Image Central!

Reply #33 –
New update:

I was never really happy with the turbo Flux quality; it takes forever and the quality for turbo is quite frankly not anywhere near as good as the pro version.
However, I now found some modified models that give really amazing results very similar to the pro model.  It takes much longer
than any other model taking over 2 minutes to create a single image, but the results are excellent.

So you now have the option to create flux.pro quality images if you don't mind waiting.  I also increased the allowed prompt length
for Flux to 4000 characters as it does understand much longer prompts.

The options are for 20 and 25 steps.  I also included a version for text in the image and a version without text in the image.
If you're not trying to add text to the image, use the [NO TEXT] options as the model uses the T5XXL 8fp leaving more room for a bigger model.
If you're trying to add text to the image, then use the [TEXT] options as they have the better T5XXL 16fp text encoder.

Some examples at 20 steps:



Re: Welcome to AI Image Central!

Reply #34 –
New update:

- Added a Stable Diffusion 3.5 Large model with the same resolutions as Flux (960x544/1920x1088)
- Added the options for 20 or 25 steps. 
- 20 steps take 90-120 seconds depending on if the model is already loaded or not.

Some examples:



Re: Welcome to AI Image Central!

Reply #35 –
I am seeing people typing in other languages on the main page in English, so I decided to auto detect the language on the English pages
as well so that you can basically type the prompt in any language and it will be translated to English.  I was having issues with the
charmap because of people typing in weird characters that are not part of the English language causing the prompt to fail.  Doing
this also solves that issue.  It really is amazing how difficult they make it to fix issues with Unicode characters!

Re: Welcome to AI Image Central!

Reply #36 –
New update:

I added progress indicators now for both GPU's so you can see the progress or see if it's busy.  I added to all models and all languages!  (18 PHP files!)
I fixed the generated image output so it's much faster to update the image.  Before it was hit or miss as far as how long it took.
And I added a colorful background behind the title area.

Re: Welcome to AI Image Central!

Reply #37 –
New update:

I decided to change the menu to make it look newer.  Finished updating all languages.
Let me know what you think.

Re: Welcome to AI Image Central!

Reply #38 –
New update:

- Made nicer buttons for Previous/Next.
- Made nicer buttons for 'Face Fix', 'Upscale X2', and 'Remove BG'.
- Moved 'Face Fix', 'Upscale X2', and 'Remove BG' to the full screen image view. (Click on the thumbnail in Requested)
- Fixed the Info window so the text wraps and made the window bigger to display 3000+ character prompts.
- Rounded the highlights and buttons to get rid of the ugly square look.

Re: Welcome to AI Image Central!

Reply #39 –
New update:

- Finally fixed an issue with CLS (Cumulative Layer Shift) at like .32 due to the slideshow.  Found a new way to do it and now CLS is an awesome .015!
Supposed to provide a good user experience.  Anything over .1 is marked as needs improvement.  Anything over .25 is considered
an error.  Hopefully now the site gets ranked better.

-Moved the sampler to the other side with the scheduler so the model is listed in the first dropdown menu now.
- Modified the page layout to display the page better and fit more in a smaller space like a tablet with a lower resolution.
- Changed the textbox's to better text input methods.
- Modified top right model links to now apply the full path from where you're at when switching models (Click the title image to go to the main page).

Screen capture of the test results:
 

Re: Welcome to AI Image Central!

Reply #40 –
New update

- Fixed placement of <title> <body> and other <meta> tags to enhance search results.
- Added latest changes to all English, Spanish, and French files.  Still 3 languages to go.
- Changed the titles and descriptions to include either the model if on the main page or category and page number.



Re: Welcome to AI Image Central!

Reply #41 –
New update:

- Added menu bar with full menu to prompt search, image search, about, and rules.
- Modified the searches so that the results appear on the same page.  Looks much better and you can see previous results.
- Added the ability to download the original PNG file by clicking the thumbnail preview on the generation page.

Re: Welcome to AI Image Central!

Reply #42 –
New update:

- Added the new ShuttleAI 3.1 Flux model to the Flux tab. Option for 4, 6, or 8 steps which is all it needs.  Pretty fast compared to the other Flux models.
Once loaded, it's only 35, 45, or 55 seconds for 4, 6, or 8 steps.  I can only load the fp8 model as the fp16 model is 23.8gb leaving no room for the
encoders.  Results are impressive for the limited number of steps.

Re: Welcome to AI Image Central!

Reply #43 –
New update:

- Finally finished updating all the other languages to the latest format.
- Changed back to textbox from the new method I was trying.  These boxes work better and can be expanded if desired.
   If you use the back button your prompt is also still in the box unlike the other method.

Re: Welcome to AI Image Central!

Reply #44 –
New update:

I have created a new feature but have not decided where to put it yet.  The new feature is Flux Fill which is a really
good inpainting model.  It takes like 4 minutes to run and 24gb VRAM but the results are very good.

The purpose of such a model is to modify an existing image where you like the image except for some flaw.
You can use the Flux Fill model to mask the part of the image you don't like and then create a prompt to modify that area.
For example, you have an image of someone wearing a shirt you really hate. You can simply use the mouse to draw
over the shirt area, then type in the prompt 'white tank top' or whatever and then wait about 4 minutes for the results
in the Requested category.

The current link is: https://aiimagecentral.com/flux-fill.html

Here is an example where I changed the eye color to red..
 
AI Image Central - Welcome to our forum!