Tested FLUX 3 Klein Locally: Has AI Image Generation Plateaued?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Tested FLUX 3 Klein Locally: Has AI Image Generation Plateaued? · Channel: Fahd Mirza · 12:04 · Watch on YouTube

Topics: AI developer tools, AI content generation, Local and open-source models, Multimodal models

Summary

Fahad Mirza installs and tests Black Forest Lab’s Flux 3 image model locally using ComfyUI, examining text-to-image generation, text rendering, multi-image references, and image editing. The model often follows prompts well and produces convincing lighting, reflections, and environmental details, but results vary: hands, realism, and some complex edits remain weak. He also describes the model’s features, VRAM use, and license considerations. His overall conclusion is that Flux 3 does not represent a major breakthrough and that image generation may have plateaued.

Key points

The video demonstrates a local Flux 3 workflow in ComfyUI and says the workflow will be shared on GitHub.
The opening flamenco dancer image follows many prompt details, but has awkward hands and an unnatural-looking torso.
Flux 3 is described as a 5.6-billion-parameter open-weights model for text-to-image, editing, and multi-reference generation with up to 10 reference images.
The distilled version uses 16 steps without guidance and GLM 3 4B as its text encoder; the model can also use bounding boxes and render text.
The presenter reports VRAM consumption of 26,418 MB, or under 27 GB, using full precision.
The Arabic and English honey-shop sign is mostly rendered correctly, though small jar labels are unreadable.
Multi-image tests preserve some subject details, while the presenter notes a somewhat plasticky appearance in parts of the results.
Editing succeeds at changing hair color but fails to correctly reposition a seatbelt; the toast-and-jam edit looks acceptable but not fully natural.
The presenter considers the Golden Temple and South Kalimantan market images among the stronger results, while the Bavaria scene looks plasticky.
He concludes that Flux 3 shows no earth-shattering innovation and suggests text-to-image and image-to-image generation may have plateaued.

Chapters

Start	Chapter	Summary
00:02	Introduction and first image test	The presenter introduces Flux 3 and ComfyUI, then evaluates a flamenco dancer image that follows the prompt but has anatomical flaws.
02:06	Model features, license, and VRAM	He outlines Flux 3’s capabilities and license considerations, and reports its local VRAM consumption.
04:11	Text rendering and multi-image tests	The tests cover an Arabic and English shop sign and using reference images to place people in a new scene.
07:21	Image editing tests	The presenter tests edits to toast, a seatbelt, and hair, with mixed results in edit accuracy and realism.
09:25	Scene generation results	He reviews images of the Golden Temple, a smoggy river market, and cows returning from the Alps, noting both strengths and shortcomings.
11:01	Mogadishu image and final verdict	After reviewing a beach scene in Mogadishu, the presenter says Flux 3 does not feel like a groundbreaking advance.

Quotes

it has followed the prompt well, which has improved, but it is not perfect by any means. — 01:34

text might not be correct in this version. — 03:08

the hair detail is quite good. — 08:54

I don’t see any earth-shattering, groundbreaking innovation in this new Flux 3. — 12:04

Concepts and entities

FLUX 3 Klein, FLUX 2 Klein, ComfyUI, GLM-3-4B, Black Forest Labs, Fahd Mirza, Vast Compute, geração com múltiplas imagens de referência, edição de imagens

Full transcript (original language)

00:02 This image has been generated by a flux client model from Black Forest Lab, their newest flux 3 model. In this video, not only I’m going to show you how you can get it installed locally, we are going to test it out thoroughly. This is Fahad Mirza and I welcome you to the channel. I’m also going to unpack how this model is different from its flux 2 version and I’m also going to share how you can get installed comfy UI, which is the tool

00:33 I’m going to use. For detailed version, just watch this comfy UI tutorial on my channel. So, let’s get a right into this. The prompt which I used was this as you can see here. I asked for a flamenco dancer in a whitewashed Seville courtyard at golden hour, mid-turn in a crimson ruffled dress with a black lace mantilla, a red carnation and a shallow depth of field. And you can easily see the result in

01:04 this image which was generated in this comfy UI workflow. Don’t worry, I’m also going to give you the link to this workflow in my GitHub repo. But look at this image now. You see, this is what has come out. The courtyard, bougainvillea, blue tiles and terracotta pots match and the warm light and soft background look right. The dress has layered ruffles and the skirt shows the spin and the mantilla and carnation are where I put them.

01:34 But there are various issues. Raised hands look quite awkward. The thumb bends oddly. Still, you know, after all these years, these models are unable to get the hand right or arms right. Um the other hand is also half hidden and the you see, if you look at this torso, it also doesn’t look that, you know, natural. So, it has followed the prompt well, which has improved, but it is not perfect by any means.

02:06 Now, if you look at this workflow, it is quite simple. One of the differences with the previous version Flux 2 client is that the Flux 2 was quite fast. This one is not as fast as I was hoping, but still that is fine as long as it generates good images. The Flux 3 image client is new open weights model from Black Forest Lab, but check out their license, too. It has got 5.6 billion parameters in the image

02:38 transformer and does text-to-image, image editing, and multi-reference generation in one model with up to 10 reference images, and we will also try it out. The distilled version needs only 16 steps, which is fast without any guidance, and lo and behold, it uses GLM 3 4 billion as its text encoder. You can also place elements with bounding boxes in the prompt. It can render text inside images, which we will test out. But, the thing is that

03:08 very surprisingly, Black Forest Lab has said that text might not be correct in this version. So, last bit on the license, it allows non-commercial use plus commercial output for qualifying user under $5 million in annual revenue. So, please do check it out if you’re looking to use it commercially. Let’s now do a few more tests. Before that, let me quickly show you my VRAM consumption. These are some of the workflows. The model is loaded, and this is my VRAM

03:38 consumption under 27 gig of VRAM, which is a bit less than the previous version, and I’m using it in full precision. And if you’re looking to rent a GPU on very good price, you can find the link to Vast Compute in video’s description. Also, please subscribe to this newsletter at fawadmirza.substack.com. And if you want to help out the channel, please become a member. For that, all you need to do is to just go to the homepage and become a member on in the first pinned comment. Okay, so now let’s

04:11 try out few more stuff with this model. And sorry, I forgot. People have been telling me to actually pronounce, you know, announce how much VRAM it is consuming. So, it is 26418 you know, MB or around you can say that under 27 gig of VRAM. And here is another result. Let me show you in full screen. So, this is from Yemen’s Sana’a city. A honey shop is signs in Arabic and English. And we are

04:42 testing whether the model renders short text in both scripts. The model has the sign got right. Honey is crisp and correct. And the Arabic below it is also okay, but I’m not sure if you’re an Arabic speaker, let me know, please. The brick tower, while, you know, the whole trim, stained glass window, and jars all match the prompt. But there are few tiny labels on the jar which are quite unreadable ones, and that makes sense given the distance of image. But you see,

05:13 the traditional attire also looks pretty good. The thing which is impressing me is the way, you know, the sunlight and the shadows and this whole stuff, you know, at the backdrop drop of this shop looks amazingly good. And look at this pattern at the top. It is quite a good result. And let me quickly show you the prompt which I used. This is the prompt. And now this is a test for multi-image. We have given this model few photos as

05:44 you can see on the left where there are um a few images which I already have inputted and then this is the prompt, a very short one, which has been upsampled. From there, this is what it has generated. So, you see, now it has built this wooden lookout, kept her hat, yellow dress, and long hair, and matched the light to the foggy orange hills. Um there are a few things which you can nitpick, but I think overall, it is not a bad image. Maybe a bit plasticky at

06:16 the foreground, this side, and also this side, but other than that, uh not a bad result. Now, for the next test, what I have done, I have just replaced that woman with my own image. And let’s see if model has been able to put me in front of that autumn forest while keeping my face and other stuff intact and removing the hat. You see, it has kept it. Everything looks good. It is even better than before, not, you know, of course, maybe because I am in that image, plus

06:48 look at these trees, autumn trees, they look a bit more realistic. There is still that plasticky look, but maybe it is trying to put that fog on the trees, but other than that, this is better. Plus, the innovation here is that it has also put a blue jeans on me, um which was not present in the original image. So, this one is quite good in terms of image editing or multi-image rendering. Plus, you can also say it is sort of editing, too. Let’s now check out image editing.

07:21 And before that, this was the prompt which I used for my previous one. Next up, let me show you how you can do the image editing. I am giving it this photo of a toast and the prompt is to spread strawberry jam on the toast and add a few fresh blueberries on the plate. Let’s see what it generates. And there you go. So, let me show you here. Nice. I’m not sure what exactly uh is behind, but that is a pattern of

07:51 that mat. But other than that, I think looks pretty good. What do you think? Um this thing, you know, the blueberries, they just look quite harmonious, bit uh AI generated one, but overall the toast texture and this jam, that looks okay. But it is by no means a world-class image in my humble opinion. The reason being it is not uh net you

08:21 know, totally natural-looking. Still a long way to go. I think, you know, the previous model was slightly better than this one in terms of realism. Okay, next step for image editing, I’m going to give it this AI generated image where I’m asking it to fix the seatbelt and put it properly on the shoulder so that it is in the legal range. And I already gave that prompt to a model, very very detailed one. And then this is what model has generated where it failed

08:54 miserably. As you can see, it was unable to edit the image to fix the seatbelt properly. Let me try a simple one to change the maybe hair color. And there you go. So, the hair have been changed. So, this is the original and this is now the edited one where model has done really really good. If you look at rest of the detail in the image, that is intact and the hair detail is quite good. It is a bit frizzy, which is quite

09:25 natural. I think model has done very very well here. And next up, with all due respect, I have tested this water reflection, the Harmandir Sahib, as you can see the prompt here, the beautiful beautiful Golden Temple in Amritsar, Punjab in India. And this is where it has created as you just saw. So, you see the gold detail, the water reflection, and the small devotees, they all look really good, the

09:55 marble inlay and mirror light reflection is quite convincing. So, the whole image is quite I would say sacred and soft, but if you are from that part of the world, if you have seen it, please also share your thoughts. What do you think about it? Yes, the birds, they look out of size and not a perfect image generation, but still it has tried. I think this is the best image this model has generated so far. I tested the smog at atmosphere and mass faces in a

10:27 busy river market in South Kalimantan in Indonesia. This is a busy river market, the orange haze, dim sun, smoke plume, and mass look quite convincing and real. All the stilt houses and background figures are also quite nice, very real. Also see on the side, this is the best one, and really my heart goes out to all these people in that region who are affected by this smog and dry season at the moment. And this is the prompt which I used.

11:01 Next up, I’m taking you to Bavaria, and this is where the the cows are returning from the Alps, and this is the image which it has generated. Um you know what? I think this one is not really there. There’s too much happening, especially in terms of plants. It has tried too hard. Looks also quite plasticky. Um Some of it is okay, but I think the realism is not there. What do you think? But looks pretty interesting concept, by the way.

11:32 And finally, we are to Mogadishu in Somalia in Africa. This is the Lido beach in Mogadishu at golden hour and basically, we are testing some colorful fabrics, a lively crowd, and turquoise water with sun satellite on sand and skin. So, overall, a good image of family sitting having a picnic. But, a detail is you see there’s just same kind of food here. Uh the beach is okay. Not bad. The reflections are really good. This is one thing which it

12:04 this model has improved. There are some minarets in the background. And some footsteps, so not a bad image. But, if you allow me to be honest, I think it seems that this text-to-image, image-to-image generation has really plateaued now. I don’t see any earth-shattering, groundbreaking innovation in this new Flux 3. But, let me know your thoughts. What do you think about it? Again, thank you for all the support.