This has been a very informative discourse. Love the relaxed style. Learnt a lot. One thing still not got an answer for ; how to assign weights to a few -sref images in the same prompt. I get images from Pinterest add to the bottom of the prompt window, but then I’m stuck, cos they aren’t part of the prompt text so I can’t :: sw them.
Thanks man! This is definitely something I hope they change. You can do this on discord, but not the site (which is annoying). Also weird how you can do with P codes and SREF codes, but not actual SREFs. - RF
Midjourney needs to add "Style Model Icons" (Realism, Anime, CyberPunk, Illustration, Custom Styles, Colors, etc....) into their UI. Adding and using "- -sref - - sw" is too cumbersome.
It can be. I do like the diversity in what you get with the weighting. It outputs more unique images because of it but I totally understand where you’re coming from. -DB
OK. So I need to have my notebook every time I come back here. Gems. 1. permutations on style weight 2.The MOODBOARD!! 3. using the --sref codes for consistent branding . The last 10 minutes just a masterclass.
Hey, love the tip about exploring srefs with mood boards! I am just wondering what your stylization and other settings are with that to get the best results.
@@MidjourneyFastHours Thank You so much for the details. I have kind of settled on --s100 so that it gives me breadboards at all. Otherwise, I usually stay within your parameters for my creations. 💞
How specifically did you create your mood boards? Did you just generate the images separately and put them together in photoshop, or did mj create the board entirely?
MJ created the entire mood board. Try this prompt: multi-panel mood board, split into multiple images, including lifestyle images and color swatches --ar 5:6 --sref [code] --v 6.1 --q 2 It will generate multiple images in one shot.
Modjourney has two kinds of stylizations: "personalization" and "reference". The Midjourney documentation confusingly uses one word, "style", in both contexts, which is very confusing. REFERENCE With reference, you make an image that looks more like one or more specific images. With "--sref URL/CODE" you can pick WHICH reference to apply. The URL can be one or more image URLs. The CODE is a specific "stylization seed", kind of like a shorthand for an existing image URL that you don't see. With the "--sw 0...1000" parameter, you control how MUCH of the reference to apply (0 means no reference, 1000 means max reference). PERSONALIZATION With personalization, you can make an image that looks more like the images preferred by particular viewers. With "--p CODE" you can pick WHICH personalization to apply (CODE is a special "time+person" stamp of preferences of a particular user). With "--p" but without supplying a CODE, you pick your own current personalization. Without --p, you use the default personalization. With the "--s 0...1000" parameter, you control how MUCH personalization to apply (0 means no personalization, 1000 means max personalization). With --style raw. Think of "--s" as actually being "pw" = personalization weight. "--s 0" still adds certain default personalization. To make the default personalization even less pronounced, add "--style raw". REMEMBER 1. --sw is the weight of the --sref. 2. --s is the weight of --p (or of the default personalization). 3. Reference pulls an image in its own direction, and personalization pulls an image in its own direction. These directions are co-dependent. 4. The best way to test the full effect of --sref is to use sref with "--style raw --s 0 --sw 1000" (and without personalization). That'll apply the full weight of the reference and will not apply any personalization.
This is all brilliant thank you so much, one thing i do not understand is if i have an image i really like and i want to mimic the style and the composition of that image but change the colours how do i do that? For example I have a portrait of a person and I want to keep the composition the same but change the person and the clothes?
Thank you so much for all the nuggets!!! They neeeeeeeeed to split out colour ref, posing, layout, style (how it's made like what type of painting, how the brush strokes are etc). Would save so many hours!
@@MidjourneyFastHours ok here goes. Keep in mind it's not possible to share an image so I'm going to try my best and use an example of a rose and a room and describe so you'll have to imagine the pictures. So basically at the moment "style" takes into account the object, colour, layout, posing, art style, etc. There's a lot in there which means it's hard to control each element if you are trying to make something specifically that you have in mind. Each of the following can have text to modify and also weighting for each image/text section (let's simplify and say everything has 1-100 like Character) with how strictly MJ should conform to the images/prompts. So what I suggest is split it into the following. 1. Object(s) (oref) The actual object or as close to it as you have. Include the ability to have multiple objects each with their own reference. This may be a character or item in the image. 2. Art style (aref) This is how the image created, it is a photo, vector, Illustration, painting? Examples: - a detailed watercolor illustration with defined edges, slight watercolor paper texture, some subtle colored linework but no black outlines, with wet-on-wet techniques. - photo: canon 1D, f2.8mm, 100mm, macro lens, 400 iso, 3. Layout (lref) The layout of the delivered art. Keep in mind that this could be of a different subject, but it's looking at the layout of the general art. Examples: - extreme close up of the centre of the rose - an asset sheet with the lifecycle of a rose in different stages - a comic book style layout showing different features - a wide angle living modern living room slightly to the side giving perspective - a flat image of the a living room 4. Color Palette (cref is taken so lets say href) Image or hex codes for the primary colours used. a weighting could then determine how strongly it has to match those colors Examples: - a color palette with x number of colors - a color palette with gradients and patterns - an image - #a3ac96 #cf7b8f Props if you've read this far. Here's an example prompt: --oref1 object1.jpg::100 --oref2 object2a.jpg::50 object2b.jpg::20 --aref artstyle.jpg::60 artstyle.jpg::40 --Lref layout.jpg::20 --href #cf7b8f #a3ac96 Additional thoughts while writing this out: - Then when you're prompting you should be able to put the references in with the text prompts if you want to use those together. ie a rose flower --oref1 object1.jpg::100 #cf7b8f::20 leaves #a3ac96 - maybe having things like aref could benefit from being on a 100% total scale so you could say it's 60% style one and 40% style two I know this may seem like overkill but for those of us who use it not for "fun" (which given a lot of the accounts I've found so far is a lot of us) this would be a game changer. And maybe it's something which they could say is for their higher tiers if they want the lower tiers to focus on "fun" and the rest of us can have our cake and eat it too. Would love to hear your thoughts.
Modjourney has two kinds of stylizations: "personalization" and "reference". The Midjourney documentation confusingly uses one word, "style", in both contexts, which is very confusing. REFERENCE With reference, you make an image that looks more like one or more specific images. With "--sref URL/CODE" you can pick WHICH reference to apply. The URL can be one or more image URLs. The CODE is a specific "stylization seed", kind of like a shorthand for an existing image URL that you don't see. With the "--sw 0...1000" parameter, you control how MUCH of the reference to apply (0 means no reference, 1000 means max reference). PERSONALIZATION With personalization, you can make an image that looks more like the images preferred by particular viewers. With "--p CODE" you can pick WHICH personalization to apply (CODE is a special "time+person" stamp of preferences of a particular user). With "--p" but without supplying a CODE, you pick your own current personalization. Without --p, you use the default personalization. With the "--s 0...1000" parameter, you control how MUCH personalization to apply (0 means no personalization, 1000 means max personalization). With --style raw. Think of "--s" as actually being "pw" = personalization weight. "--s 0" still adds certain default personalization. To make the default personalization even less pronounced, add "--style raw". REMEMBER 1. --sw is the weight of the --sref. 2. --s is the weight of --p (or of the default personalization). 3. Reference pulls an image in its own direction, and personalization pulls an image in its own direction. These directions are co-dependent. 4. The best way to test the full effect of --sref is to use sref with "--style raw --s 0 --sw 1000" (and without personalization). That'll apply the full weight of the reference and will not apply any personalization.
The Baths, ruin bars in old Jewish district, Fishermans Bastion, Budapest will have you discovering something cool everywhere. Enjoyed your talk, as a retired Video editor I think organization is paramount. Any systems that can organize work over time, volume and project would be appreciated.
I've played around with Editor, but just like your example (around 32:00 timeline), it consistently adds a bunch of unusable random stuff, even with wrong perspective, wrong lighting, wrong size ... ie: I extended the exterior of a fairly basic house and it added half a vintage pick up truck to it, adding more "house" ... if you swap faces, it looks exactly like some crappy Photoshop composit from someone who doesn't know how to composit ... the angle of the face is wrong for the position of the body, the lighting is wrong, the facial expression doesn't match the scene, and so on. I watched another RU-vidr extend the realistic, photographic image of a modern looking man and it added enormous legs with baggy pants that made him look like some circus clown with elephantitis. Even extending a simple landscape, it added junk that didn't even properly attach to very basic horizon lines nor did the size of elements in the scene match up, ie: it just tosses in huge random plants that make no sense in the scene as though they've been upsized and distorted as a lazy way to fill in the pixels. Unless a user is looking for complete abstract random images, if they are trying to use Editor to extend or retain an actual style, at this stage Editor seems completely useless. I keep trying it again every few days to see if it's been improved, but so far ... it outputs unusable 'junk'.
Totally get that and have had similar issues. To be super precise and realistic you need to work through it in various steps. I've found it hard to expand significantly and get everything correct on the first shot. Keeping the scale 75%+ and expanding has been working. It's also super sensitive to the text prompt/sref/image ref depending on what's in your initial prompt. So you have to remove the items from the text prompt or image ref/sref if you don't want it to effect the output...also dropping the style weight can help there too. It's somewhat complex.
