Image generator tip
Wanna say big thank you to the creator of perchance as this site has gotten me to actually understand coding better. Moving from SD to Flux.1 Schnell has also been a major boost. I am working on a side project image Gen that also ties in the ai chat and going to make a game generator that you can play a DND type game with the ai chat bot that generates images of the characters and the scene, still not close to getting that done.
Now, biggest thing I have been seeing with people and posts about the t2I generator now is that they do not understand Flux.1 at all. Here are some tips.
-
Flux does not use negative prompting. You can add No "whatever" in the main prompt at the bottom to help Flux not generate something you dont want in the image.
-
Flux does not use weights or () or any SD type prompting. Use the KISS method. Your prompt should be simple, ie: A husky dog running in a wooded forest during the middle of winter. - that will generate a good image. You can add - real photograph taken on an Kodak disposable camera, y2k aesthetic, natural lighting and light leaks through the trees. - by adding the technical data to the end, Flux actually responds to it better. You can mention type of trees, or more background type things, just remeber the KISS method.
-
Flux does not like high SFG. I hope the dev can adjust perchance main code to allow 1/2 scale so so you can do 2.5, 5.5, etc instead of whole numbers. Flux does prefer 1-5 SFG.
-
I am not sure if perchance can actually do image steps, I have been messing with that part and it does seem to adjust the quality of the image. Under the image settings code, add a steps = [inout.steps] code line and create a simple step list. I go with 4, 8, and 12 steps. Just something to look into.
-
Using the old perchance imports for art styles and such, is actually not right. Those are setup for SD and not Flux. Using them without changing the prompt structure will always create bad images.
Seen to many posts here and on reddit where users keep complaining about the t2I generators and wanted to give some insights to help the community. Using Claude, Gronk, Chat gpt to help build prompting is decent, again they tend to over describe things. The ai chat bots you can create do a better job once you setup the persona correctly to fully understand how Flux.1 works and give the bot good guides and instructions to follow.
Thank you for your insightful tips, I'm definitely a common user of Perchance t2i and have no real knowledge about Flux or how I should create the prompts to better instruct Flux to create what I want. That said, my personal 'trial and error' experience is different than what you just stated. For example, still today I notice visible differences when enclosing something in brackets; not sure whether Perchance makes any difference when using () or ((())), but still there's a difference, especially if the bracketed element is in the last part of the prompt. I'm not sure what the SFG is 😅. If it's what the default Perchance t2i interface calls 'Guidance Scale', then a value of 1-5 gives horrible results with realistic images; not so bad when it comes to render a painting or drawing. My prompts are quite complex, and I often use a Guidance Scale ranging from 10 to 20, to balance the adherence of the image to the prompt and the realistic output. Lastly, I've modified some of the standard Perchance styles to suit my needs, but the evergreen "Cinematic" style (and a few others) is just perfect as it is, and still gives outstanding results.
You have to make sure your prompts are concise and not too wordy. Leave out generic terms like masterpiece, bad quality, and go for technical camera prompts or photorealistic image of.... the brackets may not be the reason for the visible difference with flux. Having a good prompt structure is what works. Use Claude or chargpt to give better tips on how to setup your specific generator. Those have really made a difference in ensuring a more natural language flow in the prompting. I will test out the brackets too and see if that does help or not. But right now a guidance scale of 7 with a step of 8 for certain photoreal artstyles is working really good for me.
Here is some good photo prompts I use. I have the [input.artStyle.a] tag at the beginning of my prompt setup that then feeds into the character build for anime/fictional characters then into other prompts and have the [input.artStyle.b] tag towards the end of the prompt structure. This way flux see the key details of what to generate then sees the aspect of the artstyle. My biggest issue is with webcam, selfie, phone type cameras. Cannot get it to consistently generate real depiction. Only illustration or cartoonish images. I am still tweaking these as I go too.
Bro is justpassing 2.0