Get The BEST of Both Worlds in DALL-E 3 & Midjourney

[@mikeosterberg4611]

My theory on why faces look “fake” is that it’s not Dall-e’s fault. As a graphic designer, when I retouch images in Photoshop, I do not completely remove wrinkles and lines with automated healing brushes or other tools, but dial facial imperfections down 20-30% so the person’s character remains and you can still see skin pours. However, most people use free photo retouching tools to totally blur out their imperfections for their social-media (and online dating) posts. Thus, since Dall-e and Midjourney source their image generations from available images on the web, the most common images they will source are these horrible, blurred-out skin tone images, because that is how they see people. We can’t blame Dall-e or Midjourney. We have to blame Tinder, Okcupid, Bumble and Instagram, and the vanity of 20-somethings.

[@TokenizedAI]

Doesn't explain the overly pointy nose and Midjourney produces very natural faces. So clearly it's possible.

[@unglueme]

I don't know if that's true...but it's a thought and thanks for your perspective. I use Bing Image Creator a few hours per day since i first found out about it two and a half months ago. It's gone downhill in MOST ways since they started to have DALLE-3 and improved in maybe a few ways. But the worst thing isn't even that, but that i can't even use the word "zombie" or "foot" or just about 75 percent of words i normally use without it NOW JUST STARTING TODAY or yesterday saying it's "UNSAFE". so it's now completely unusuable due to that problem.

[@morpheus2573]

This society will be remembered by history primarily for its toxin injected duck-lips. 👀 🦆

[@mazenkilani]

Thanks for the insights, truly appreciated. In the latest Midjourney Office Hours, it felt the Midjourney team does not know, or does not want to know, about the existence of DALL-E 3. I am a paid subscriber of Midjourney, but now feel they are off-track in their priorities. I enjoy using Midjourney, but wish they can re-focus based on the community reactions to DALL-E 3; the old style of prompt 'engineering' structures and post-manipulation of images feel so archaic after using DALL-E 3.

[@CarletonTorpin]

8:28 - In GPT4, using Dall-E 3, I've gotten widescreen-aspect-ratio images by starting my prompt with "please make a landscape image of __________" and the same would work for vertical-images by asking for "portrait-oriented image of _________". I have not tried it in Bing Chat though; I've only used Chat GPT 4 with the Dall-E 3 option enabled.

[@TokenizedAI]

It seems that Bing Image Creator and ChatGPT are using different models. The quality is vastly different between both.

[@CarletonTorpin]

@@TokenizedAI I've been thoroughly impressed with GPT-4 and Dall-E 3, especially for its ability to correctly place phrases of text into images. With the natural-language prompting ability, GPT-4 + Dall-E 3 is an amazing tool for ideation.

[@ScaryStoriesNYC]

I agree with you about MidJourney's censorship. Sometimes it censors me if I request specific clothing on a female character to match previous images it generated for me, it does it but then makes the page blue so I can't do the face swap! So then I have to select a small part of the background and do a "Vary region" overwriting the entire prompt with something inoffensive like "a light bulb." Then I choose one of the options where it did NOTHING to change the image (there's always one like that) and I can do the face-swap on THAT because it's not describing women's clothing. BTW you can talk about men's clothing all day without getting censored, MidJourney is very sexist against women though, it's drastic.

[@PortugueseLion]

I am not sure about your last remark based. MidJourney is as prudish on men's clothing and body's as much as on woman's. It seems to be not sexists in the way that in this respect it treats man and woman the same, what is out of line with society. At least mine. For instance a bare-chested man is as unacceptable for MJ as a bare-chested woman.

[@jeremysomers2239]

brilliant as always, consitently coming up with KEY workflows, thank you, Christian!

[@PreciousEyeballs]

3:32 I don't think Bing Image Creator will accept the word "blood" as well. It's been rejecting even the most harmless prompts lately.

[@TokenizedAI]

But it accepts "dark crimson red paint/liquid". ChatGPT doesn't because it understands the context and doesn't like it.

[@PreciousEyeballs]

@@TokenizedAI I prompted someone wearing a red outfit, it won't accept until I removed the word "red." 🤷

[@TokenizedAI]

How dirty of you! 🤣🤪

[@anav.8719]

Christian, you hands-down have the best channel on KZbin that deals with this stuff! Tanks so much for taking the time to do this.

[@KolTregaskes]

I won't be cancelling MJ. v6 sounds like it will be bringing DALL-E 3 like prompt translation and there are various new features coming before then (+3D and some form of video before the end of the year or early next) so MJ is still the best image generator IMHO. Also the quality of images is a lot better in MJ. I cannot get close to MJ's photo-reaslim in DALL-E 3. But for now this workflow is really good thank you. :-)

[@jasonwiley3617]

Told you great prompting but still has a way to go before beating Midjourney out but it is a great first step to improve prompting.

[@barto.streetart]

Excellent tip and study-cases! Thanks a lot!

[@LouisGedo]

1:33 Yup, that's me by the end of December 2023 if MJ doesn't come out with a radically upgraded version which actually completes with Dall-E 3 in quality and coherence.

[@CarletonTorpin]

Great stuff. My anecdotal experience with Dall-E 3 is that it will "not" make images for prompts that include the word "Macbook" , but it will happily make a photoreal image of "laptop" with a picture-perfect Apple logo on it. In short, it can make an image of anything, but it has guardrails up for certain words and phrases (possibly to "protect' IP of these businesses being mentioned).

[@TokenizedAI]

You have to avoid all references to brand names and products.

[@CarletonTorpin]

@@TokenizedAI Your videos, along with a few others on KZbin, provide the best user-manual & discourse for these nascent technologies. Thanks for facilitating useful discussion in the image-generation world. :)

[@TokenizedAI]

I aim to please 😁

[@TheoreticallyMedia]

Great video, Christian! I still think the real powerhouse is going to be when Dall-E 3 is fully integrated into ChatGpt, provided we can edit/modify our initial outputs via text prompting. Basic things like in and outpainting, but doing so in a conversational way. The announcement video kind of implies that this will be the case, but... Having experimented a bit with ChatGPT Vision, I'm actually concerned this won't be the case. Despite the fact that GPT (at least for me) isn't capable of generating images, I still gave it a shot and it did the whole lying thing, providing me with broken links. I decided to push it and lied back, asking it to change aspects of the nonexistent image, and GPT responded that it couldn't do so. That it was only capable of generating one image with no modifications. Now, that might just be another lie on its part, considering it isn't the updated model...but, it does have me a bit wary. If Dall-E 3 is going to be a competitor to Midjourney, it needs to have some editing and variation features.

[@TokenizedAI]

I have DALL-E 3 access in ChatGPT now and actually recorded a nice comparison video with Igor (The AI Advantage) which we'll release this week. DALL-E 3 via ChatGPT is superior in image quality but it messes around with your prompt because it doesn't take your prompt verbatim. Instead, it rephrases your prompt with multiple variations, subsequently also changing your intent sometimes. That being said, the conversational nature of making prompt adjustments is very intriguing and quite a bit of fun.

[@TheoreticallyMedia]

@@TokenizedAI That's awesome. Looking forward to seeing that video! The rollout has been so odd in terms of who gets it when, but I'm really happy you got access! Every morning I wake up waiting for that AI Christmas Day!

[@TokenizedAI]

@@TheoreticallyMedia I don't have access to ChatGPT Vision 😂 It's so random. An interesting tidbit is the fact that Bing Image Creator and ChatGPT don't seem to use the same DALL-E 3 model. The results are very different. And while Bing will do R-rated stuff, ChatGPT is in Care Bear land.

[@TheROZ]

I agree. The addition of Dall-E 3 access to chat gpt already allows for users to create different aspect ratio images rather than bing image creators 1024x1024 default. Often times with character design ive noticed that bing image creator doesnt allow for multiple poses cleanly but the extended options via aspect ratio in chat gpt has so far sufficed my characters.

[@morpheus2573]

The image of you trying various techniques to catch out an AI who is trying to lie to you is equal parts hysterical and cray-cray... in a good way. 😜

[@ArtofGhori]

Great tip, thanks

[@timtim2500]

Great video as usual, Christian... can you pls share the prompt you used for the THUMBNAIL of this video? I know you added the logos in photoshop, but talking about the rest of the image and styles etc... Thanks again

[@TokenizedAI]

I actually can't because it's a result of a longer conversation with ChatGPT. It's not a single prompt per se. 😉

[@erikaol1091]

Supposedly, midjourney v6 will have a better prompt understanding. I wonder how good it is going to be.

[@TokenizedAI]

"Supposedly" seems odd unless this is based on a statement from Midjourney itself.

[@beamsgr]

I really liked your comment about the faces looking slightly fake. Typically with new visual technologies like this our eyes are not yet trained to see the artificiality of the images. As we become increasingly familiar with the technology, we notice more and more of the tell-tale signs of artificiality that we previously missed. I expect that in a few years as our eyes become more sophisticated, then the images we see generated by AI today that seem so realistic, will look obviously fake to us.

[@astrodas1910]

In a few years, images generated by AI will look more realistic than photos taken by cameras.

[@jrikkerink]

Could you try to do something similair with detailed architecture? Couldn't get this right at this point. Busy for a while creating photorealistic festival stage designs which is pretty difficult in Midjourney because of the many specific details, but now with Dall-E 3 it's a big jump forward! A downside is the 1024 x 1024 resolution and no chance to zoom or change scale. Trying to get these Dall-E creations into midjourney and zoom out to get a detailed full view in 16:9, but sadly Midjourney can't get these designs equivalent to the original..

[@The-Spondy-School]

Although I appreciated the coherence, I ran into similar limitations when when using Dall-E3 in Bing. It didn't occur to me at the time to move the Bing pics into MJ. Helpful tip -- TX. BTW -- I went to Chat GPT 4 this morning and looked to see if they have yet integrated Dall E 3 into their program, and did not see it. Do you / anyone have a clue when it will show up there?

[@TokenizedAI]

I have it since Thursday. The rollout is completely random.

[@bennguyen1313]

Aside from the age-old question of Midjourney vs Un/Stable Diffusion , and now vs Dall-E 3 / Bing Image Creator... Any thoughts on using Fooocus/SDXL? You don't have total control (ex. CFG values) but it does seem to help in learning how some controls affect the image? Also, with so many tools (ideogram ai , BlueWillow, Playground ai, Leonardo ai, ControlNet Tile,Realistic Vision V5.1, After Detailer, Automatic1111, DreamBooth), Any thoughts on mixing tools together? For example, using LyCORIS (LoCon and LoHa) with LoRA?

[@TokenizedAI]

Sure, I mix tools all the time, especially when one tool serves as a fix for some specific thing. But let's face it. No one really wants to use a gazillion different tools. There are already far too many out there. And many of those that you mention only have users because they're free (for now). If people aren't paying for it (SD users are paying for hardware and electricity) then it isn't gonna stick around in the long term anyway. The market is going to consolidate on 2-3 heavy hitters. And the others need to find ways to differentiate themselves into a specific niche or unique use cases.

[@AliceAvery]

Great video and great use of both Dall-e 3 and Midjourney, love to know Christian @TokenizedAI is it possible to in midjourney to take my created images from dall-e-3 and have midjourney create the image again basically so I can upscale it as 1024 x 1024 is too small. Would there be a prompt to have it follow the image uploaded to as close as the original as possible, to have the ability to add outpainting to the image and make it a portrait would be awesome then I could upscale it with gigapixel which I was going to get at black friday, very curious to know about Midjourneys upscaling coming soon though, maybe I won't need gigapixel.

[@TokenizedAI]

You can't replicate just any external image in Midjourney via an image reference. It's neither technically possible right now nor is it something MJ wants to enable. Just consider all the potential negative consequences of something like that.

[@AliceAvery]

@@TokenizedAI Thanks Christian, yes you are right, didn't think some would would abuse this, so would you have any other suggestions on upscaling something from dall-e 3 or best to wait to see if they offer something in the future.

[@TokenizedAI]

If all you want to do is upscale the image, then either use one of the many free upscalers out there or figure out how to install Stable Diffusion locally because the latter gives tons of upscaling options at absolutely no cost.

[@AliceAvery]

Thanks Christian@@TokenizedAI yeah tried a couple of them yesterday and didn't get great results, maybe look for more of them. yeah would love to try stable diffusion but sadly my PC doesn't have the high spec for it. Thanks anyway, I'll maybe see if Midjourney with the same prompt and a reference image would maybe give me something even better like you showed in your video.

[@conhuir]

I got access a few days ago and I've already canceled my Midjourney sub. So much easier to work with.

[@TokenizedAI]

Not an option for a lot of people like myself. MJ's image quality is still superior and DALL-E 3 is still too limited in advanced editing. But I understand that for many people it's enough.

[@Mooooty]

Midjourney's aesthetics and image quality...are so much higher than DALL-E 3...not even a comparison.

[@TokenizedAI]

Are you using Bing Image Creator or ChatGPT. There is a considerable difference. They're not the same DALL-E 3, as far as I can tell.

[@5timesm]

Interesting combination workflow…. 👍🏻

[@ultimategolfarchives4746]

Midjourney has a premade idea, anx will take some of you idea into consideratiom. Dall-e sets forth with an unscribed tableau, always attentive to your imaginative cues. Dall-e doesn't produce the same realistic results as Midjourney because it follows your specific requests. Refine your prompt further, and you'll achieve outcomes surpassing Midjourney's quality. It just necessitates more precise prompting. Here, you've opted for familiar visuals. But when Dall-e is provided with more avant-garde and complex composition, it truly demonstrates its superior capabilities compared to Midjourney. Great video as always 👍👌

[@TokenizedAI]

Having tested DALL-E 3 quite extensively now, I can say that it's incredibly useful for a lot of things. And it's coherence is unmatched right now. But I strongly disagree on the quality. It's very good, yes, but it's still not on MJ's level. The mere fact that you cannot make fine-tuned adjustments to an existing image (i.e. via parameters) without producing something very different is already a huge limitation. Especially with anything that requires more consistency. It's still up to randomness. The initial randomness is just more controlled. Regarding precise prompting, the very same could be said of Midjourney. It's just an entirely different prompting style.

[@ultimategolfarchives4746]

@@TokenizedAI I agree with you. Both have their strength and weakness, and both complete well together. Dall-E don't offer much tool to refine yet and the seeds in GPT don't work for some reasons. Hopefully they will probably fix that soon 🤞🤞🤞. They will surely come up with tools eventually. They just released the model, so we have to give them a chance :-). Since the seeds don't works, it offer one way to fine tune images. In GPT, write your prompt, and tell gpt to keep the prompt as it is for every images. You'll get 4 identical images (since they all have the same seed). Then, you can copy/paste 4 times the same prompt and remove/add/change some keywords. You'll end up with results close to the original image. P.S. I wouldn't come back to Midjourney's prompting system even if the quality is higher 🙉🙈.

[@lastchildz8567]

This is correct but it will be suitable for limited image generation. If they plan to stay like that. Let's not forget that Midjourney has a more advanced edition (pan, custom zoom, inpainting etc...)

[@_gabrielln.dop_]

I have a question about your course -- I often find these courses are extremely basic and aimed at beginners, thankfully a lot of them have return policies within a certain time frame. Do you offer something similar in case the course is not what I'm looking for?

[@TokenizedAI]

I don't have a structural refund policy simply because it makes digital products an easy thing to abuse. However, I can assure you that I've never rejected a refund request that was accompanied by a reasonable justification and proof that someone actually put in the work. I strongly recommend you read through the full sales page because it addresses the most common concerns that might lead to disappointment (e.g. incorrect assumptions). That being said, after more than 260 students of varying degrees of experience, not a single one has expressed any form of dissatisfaction, let alone asked for a refund. I think that speaks for itself. I sincerely believe that's because the course is very well-balanced and I actually care about the quality of the products I put out. 🙂

Thanks !!

[@vincentburroughs7451]

dalle 3 has been on pc version for over 10 days now

[@TokenizedAI]

There is no "PC version" of DALl-E 3. You mean Bing Image Creator. Please read up on how these rollouts work.

[@Brismo7]

I feel like DALL-E 3 is using some of the code that Midjourney is using. They look so similar.

[@charleswoodruff9013]

The viking is overlooking the fjord, not overseeing.

[@TokenizedAI]

Semantics 🙄🤦🏻‍♂️

[@serenditymuse]

Can't bear to use anything Microsoft so I will wait for another path. I am linux and Android only. I guess I can try it on Android since I need Edge.

[@TokenizedAI]

I don't understand what you are saying. All of these are web-based and none of them require an Edge browser.

[@xuantung191]

Great

[@tomagoredo]

midjourney PG13 censorship is insulting and complete rubbish

[@TokenizedAI]

Wait till you see what ChatGPT does. It's even worse.

[@darkregentt]

Overrestrictive tool, its fun but frustrating.

[@darkregentt]

I just try "Girl sitting in a dark alley" and its restricted... Its AI image generator to create whats on your mind (with some limitations) but in other hand its not, because 90% of what I want to create is restricted

[@TokenizedAI]

This sounds like DALL-E 3 via ChatGPT. You could probably create that easily in Bing Image Creator.

[@darkregentt]

No I am using bing... Now I wrote just random "Army soldier with a pink mask sitting in a bus, everybody is laughing around him, 80mm film" and blocked...What is offensive on this? :D I dont get how it decide what is offensive and what is not if even this dont work@@TokenizedAI

[@Edino_Chattino]

I figured out that the "problem"word here is "girl". I think it assumes you're trying to make children do nasty stuff. I found the way around it o use young female, or small female. BTW, I was trying to make a catgirl, and the only way around was to ask for a cat woman (not catwoman as it gave me the character).@@darkregentt

[@LouisGedo]

👋

[@vincentburroughs7451]

and the video says it is 2 hours old, so does this guy not have a 20 dollar chat gpt account?

[@TokenizedAI]

Two things about "this guy": 1. I record these a few days in advance. 2. DALL-E 3 for ChatGPT hasn't been rolled out to everyone. I have it since yesterday and others still don't have it. You know what they say about making assumptions....

[@BiteSizeNarratives]

The advantage of Dalle 3 is that it can produce accurate images from difficult prompts, but for photography style it is still worse than Leonardo and Midjourney

[@TokenizedAI]

While that's true for Bing Image Creator, the output from ChatGPT is levels beyond Bing Image Creator. It has other limitations but it's pretty clear that they both do not use the same model. Don't ask me why 🤷🏻‍♂️

[@harrygoldhagen2732]

Excellent, Chris! Thanks so much. I am getting such a great education from you and Tim at @TheoreticallyMedia and a few others who keep up with all this fascinating technology.