Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

FLUX.2: Frontier Visual Intelligence

bfl.ai

Post reply on HN
Print view View original

111–120 of 124 posts

Re: FLUX.2: Frontier Visual Intelligence

#111

Post by kevin_thibedeau »

loudmax wrote: post #63
cubefox wrote: post #61

Earlier quoted context omitted.

I think it's more likely this is just a niche that Midjourney has occupied.

If Midjourney is a niche, then what is the broader market for AI image generation? Porn, obviously, though if you look at what's popular on civitai.com, a lot of it isn't photo-realistic. That might change as photo-realistic models are fully out of the uncanny valley. Presumably personalized advertising, but this isn't something we've seen much of yet. Maybe this is about to explode into the mainstream. Perhaps stock…

> what is the broader market for AI image generation?

Replace commercial stock imagery. My local Home Depot has a banner by one of the cash registers with an AI house replete with mismatched trim and weird structural design but it's passable at a glance.

Re: FLUX.2: Frontier Visual Intelligence

#112

Post by woolion »

vunderba wrote: post #104
woolion wrote: post #98

Earlier quoted context omitted.

The comparison are very useful but also quite limited in terms of styles. Models tend to have extremely diverse abilities in following a given style against steering to its own. It's pretty obvious that OpenAI is terrible at it -- it is known for its unmissable touch. However, for Flux it really depends on the style. They already posted at some point that they changed their training to avoid averaging different style…

The site is broken up into "Editing Comparison" and a "Generative Comparison" sections. Generative: https://genai-showdown.specr.net Editing: https://genai-showdown.specr.net/image-editing Style is mostly irrelevant for editing, since the goal is to integrate seamlessly with the existing image. The focus is on performing relatively surgical edits or modifications to existing imagery while minimizing changes to the re…

I didn't go very far with my own benchmarks because my results were just so bad. But for example, here's a line art with the instruction to color it (I can't remember the prompt, I didn't take notes).

https://woolion.art/assets/img/ai/ai_editing.webp

It's original, ChatGPT, Flux.

Still, you can see that ChatGPT just throw everything out and does not do a minimal attempt at respecting style. Flux is quite bad, but it follows the design much more (although it gets completely confused by it) that it seems that with a whole lot of work you could get something out of it.

Re: FLUX.2: Frontier Visual Intelligence

#113

Post by vunderba »

woolion wrote: post #112
vunderba wrote: post #104

Earlier quoted context omitted.

The site is broken up into "Editing Comparison" and a "Generative Comparison" sections. Generative: https://genai-showdown.specr.net Editing: https://genai-showdown.specr.net/image-editing Style is mostly irrelevant for editing, since the goal is to integrate seamlessly with the existing image. The focus is on performing relatively surgical edits or modifications to existing imagery while minimizing changes to the re…

I didn't go very far with my own benchmarks because my results were just so bad. But for example, here's a line art with the instruction to color it (I can't remember the prompt, I didn't take notes). https://woolion.art/assets/img/ai/ai_editing.webp It's original, ChatGPT, Flux. Still, you can see that ChatGPT just throw everything out and does not do a minimal attempt at respecting style. Flux is quite bad, but it…

Yeah so NOVEL style transfer without the use of a trained LoRA is, to my knowledge, still a relatively unsolved problem. Even in SOTA models like Nano Banana Pro, if you attach several images with a distinct artistic style that is outside of its training data and use a prompt such as:

"Using the attached images as stylistic references, create an image of X"

It's fall down pretty hard.

https://imgur.com/a/o3htsKn

Re: FLUX.2: Frontier Visual Intelligence

#114

Post by vunderba »

spunker540 wrote: post #108
raxxorraxor wrote: post #102

Earlier quoted context omitted.

I think the margin isn't that large to be honest. If we compare available resources and data it is quite tiny and perhaps should be larger. Also it doesn't feel solved to me at all. There is no general model, perhaps it cannot reasonably exist. I think these tests are benchmarks are smart, but they don't show the whole picture. Domain specific image generation tasks still require a domain specific models. For art pur…

Does SD1.5 suffer from resolution / coherence / complexity issues? I understand most outputs could be fine tuned for most domains, but still felt sd1.5 had a resolution ceiling, and a complexity ceiling no matter how good the fine tuning

Yeah SD 1.5 is mostly trained on datasets of resolution of 512x512. That's why you'd get crazy multi-limb goro abominations if you pushed checkpoints too much higher than 768x768 without either using a Hires Fix or Img2Img.

There's not much of a reason to use SD 1.5 over SDXL if image quality is paramount.

A lot of people (myself included) use a pipeline that involves using Flux to get the basic action / image correct, then SDXL as a refiner and finally a decent NMKD-based upscaler.

Re: FLUX.2: Frontier Visual Intelligence

#115

Post by woolion »

vunderba wrote: post #113
woolion wrote: post #112

Earlier quoted context omitted.

I didn't go very far with my own benchmarks because my results were just so bad. But for example, here's a line art with the instruction to color it (I can't remember the prompt, I didn't take notes). https://woolion.art/assets/img/ai/ai_editing.webp It's original, ChatGPT, Flux. Still, you can see that ChatGPT just throw everything out and does not do a minimal attempt at respecting style. Flux is quite bad, but it…

Yeah so NOVEL style transfer without the use of a trained LoRA is, to my knowledge, still a relatively unsolved problem. Even in SOTA models like Nano Banana Pro, if you attach several images with a distinct artistic style that is outside of its training data and use a prompt such as: "Using the attached images as stylistic references, create an image of X" It's fall down pretty hard. https://imgur.com/a/o3htsKn

I'm pretty sure that some model at least advertised that it would work. I also think your example was in the training data at some point least, but I suspect these styles are kind of pruned when the models are steered towards "aesthetically pleasing" outputs which are often used as benchmarks. Thanks for the replies, it's quite informative.

Re: FLUX.2: Frontier Visual Intelligence

#117

Post by dragonwriter »

loudmax wrote: post #63
cubefox wrote: post #61

Earlier quoted context omitted.

I think it's more likely this is just a niche that Midjourney has occupied.

If Midjourney is a niche, then what is the broader market for AI image generation? Porn, obviously, though if you look at what's popular on civitai.com, a lot of it isn't photo-realistic. That might change as photo-realistic models are fully out of the uncanny valley. Presumably personalized advertising, but this isn't something we've seen much of yet. Maybe this is about to explode into the mainstream. Perhaps stock…

> Porn, obviously, though if you look at what's popular on civitai.com, a lot of it isn't photo-realistic.

I don't have an argument to make on the main point, but Civitai has a whole lot of structural biases built into it (both intentionally and as side effects of policies that probably aren't intended to influence popularity in the way they do) that I would hesitate to use "what is popular on Civitai" as a guide to "what is attractive to (or commercially viable in) the market", either for AI imagery in general or for AI imagery in the NSFW domain specifically.

Re: FLUX.2: Frontier Visual Intelligence

#118

Post by dragonwriter »

_ache_ wrote: post #25
minimaxir wrote: post #12

Earlier quoted context omitted.

Looking at the file sizes on the open weights version ( https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/mai... ), the 24B text encoder is 48GB, the generation model itself is 64GB, which roughly tracks with it being the 32B parameters mentioned. Downloading over 100GB of model weights is a tough sell for the local-only hobbyists.

Even a 5090 can handle that. You have to use multiple GPUs. So the only option will be [klein] on a single GPU... maybe? Since we don't have much information.

> Even a 5090 can handle that. You have to use multiple GPUs.

It takes about 40GB with the fp8 version fully loaded, but ComfyUI can (at reduced speed), with enough system RAM available, partially load models in VRAM during inference and swap at need (the NVidia page linked in the BFL announcement specifically highlights NVidia working with ComfyUI to improve this existing capacity specifically to enable Flux.2) to run on systems with too little VRAM to fully load the model.

Re: FLUX.2: Frontier Visual Intelligence

#119

Post by dragonwriter »

throwaway314155 wrote: post #10
minimaxir wrote: post #4

Text encoder is Mistral-Small-3.2-24B-Instruct-2506 (which is multimodal) as opposed to the weird choice to use CLIP and T5 in the original FLUX, so that's a good start albeit kinda big for a model intended to be open weight. BFL likely should have held off the release until their Apache 2.0 distilled model was released in order to better differentiate from Nano Banana/Nano Banana Pro. The pricing structure on the Pr…

> as opposed to the weird choice to use CLIP and T5 in the original FLUX This method was used in tons of image generation models. Not saying it's superior or even a good idea, but it definitely wasn't "weird".

Considering how little (and sometimes negative) benefit it provided in most of them compared to just using the biggest encoder model and having a null prompt on the rest (not just those using the specific combination Flux.1 did, but for most of the multi-encoder models), its actually pretty weird that people kept doing it.

Re: FLUX.2: Frontier Visual Intelligence

#120

Post by vunderba »

woolion wrote: post #115
vunderba wrote: post #113

Earlier quoted context omitted.

Yeah so NOVEL style transfer without the use of a trained LoRA is, to my knowledge, still a relatively unsolved problem. Even in SOTA models like Nano Banana Pro, if you attach several images with a distinct artistic style that is outside of its training data and use a prompt such as: "Using the attached images as stylistic references, create an image of X" It's fall down pretty hard. https://imgur.com/a/o3htsKn

I'm pretty sure that some model at least advertised that it would work. I also think your example was in the training data at some point least, but I suspect these styles are kind of pruned when the models are steered towards "aesthetically pleasing" outputs which are often used as benchmarks. Thanks for the replies, it's quite informative.

Sure! So that image was pretty zoomed out, I've gone ahead and attached some of the reference images in greater detail:

https://imgur.com/a/failed-style-transfer-nb-pro-o3htsKn

Now you should be able to see that the generated image is stylistically not even close to the references (which are early works by Yoichi Kotabe). Pay careful attention to the characters.

With locally hostable models, you can try things like Reference/Shuffle ControlNets but that's not always successful either.

Post reply on HN