looks like it,, more free stuff to play with, I like it

they can try, but good luck with that! :D Qwen is the best

Unless you're talking Edit, even Flux 1 destroys it.
I would be staggered if Flux 2 wasn't two leagues above Qwen...

@yorgash Ofc I was talking about edit, offering the same kind of features. I really wonder what will be the result of this battle.

@yorgash flux1 never destroyed anything but itself. now, why would you say that? does flux have more parameters, better text encoding,better licensing? in exactly what areas is flux better than qwen-image?

@sweetmax797 Everything visually but prompt adherence.
Qwen is the only model that can't create decent looking pictures on its own.
Krea, SRPO, Flux, Chroma and WAN are all great, while Qwen is sub SDXL level at almost everything but prompt adherence (which costs it almost all variety and creativity).

@yorgash 

run this in flux and then in qwen-image ```text
Ultra realistic portrait of a menacing cyberpunk man with a deranged expression, highly detailed skin textures, pores, and metallic cybernetic enhancements under harsh lighting, vibrant color saturation with glossy reflections and subtle shadows, set against a solid bright yellow background.

Starting from the top: Spiky, voluminous hair in deep navy blue with glossy highlights and subtle gradient to lighter teal at the tips, large and wild, swept upward and backward from the forehead, occupying the upper third of the frame centrally.

Forehead shows realistic wrinkled skin with a prominent red scar vein running vertically on the right side, medium size, positioned slightly off-center right.

Large metallic visor goggles cover the eyes, angular and futuristic shape with chrome silver frames and deep blue lenses, featuring bold vibrant red X symbols glowing with neon intensity across each lens, oversized and dominating the mid-face, centered horizontally with metallic sheen and light reflections.

Ears are human-like with subtle wrinkles, medium size, positioned on sides; small black stud earring on the left ear lobe.

Nose is straight and pronounced, realistic flesh tones with shadows under the visor.

Mouth grins maniacally, wide open revealing sharp white teeth with realistic enamel shine and slight yellowing on canines, large and stretched, lower lip pinkish with texture, centered below the visor.

Chin and jawline are angular and cybernetically augmented with metallic plates in gunmetal gray, integrated seamlessly into skin with rivets and shadows.

Neck partially visible, muscular with visible veins, leading into a worn blue leather jacket collar, glossy cobalt blue with scuff marks and yellow paint splatters, large lapels folded outward, positioned at the bottom third.

Jacket features metallic armor plating in dark steel gray underneath, riveted and textured, with small circular buttons on lapels: left has a silver button, right has a red circular patch with black symbol resembling a stylized anarchy sign, medium size, positioned lower left and right on the collar.

Vantage Height: Eye-Level. Camera angle: Straight-on close-up.
```

@yorgash Qwen works great with the kind of setup that I use, but it is difficult to use I admit. Looking forward to seeing if Flux 2 no-lora can match up to peak Qwen.

@yorgash btw what do you mean by 'on its own'? Have you ever run the full model and text encoder without LightX2V 4 or 8 steps? as far as realism goes both model need to kneel in front of SDXL and perform fellatio :)

@sweetmax797 Yeah, especially considering the initial cost of the computer if you want to use locally. SDXL-based models are still beasts running on 10+years old GPUs

Never, Unfortunately, Only Flux2 pro API service can compete with Qwen Image edit2509 but We will have Qwen Image Edit 2511.

Он и промпт понимает ру

@EPIC__fail наша слоняра!!!!!!!!!!!

@EPIC__fail Due to the text encoder 

It's almost a day-old model, there will definitely be optimization attempts. I tested it on 10 GB VRAM (CPU: i7, 64 GB RAM). Surprisingly, 'it/s' was lower than Flux1 Dev, but try at least 10 generations of totally different subjects. Its speed will improve a bit.

uh was about to click dl... I will wait a bit x))

I added ReferenceLatent, and it became a few times slower, why

On my 8 GB it took 874.50s. The quality is worse.

@slonovich040403842 I need ~255 seconds

It took 440s on my 3090Ti... RIP. I sure hope it gets faster. Need to train some LoRAs too because it still has plastic look and isn't quite photo-realistic. Always feel like starting over again each new model. I'm sure it'll get there eventually though. ... Ok, I got it to 278s maybe after model was loaded and maybe having the clip use cpu? Not sure. The mistral_3_small_flux2_bf16 clip is huge but I haven't found any other that worked yet.

i have a 5090 u will give you an update soon on the time xD i will use fp8 version (q version is for......yeah i will use the regular one xD)

Пожалуйста! Как насчет modelscope.cn? Он тоже заблокирован?

Кстати, да вообще не смогу скачать, а с VPN скорость 200 Кб/с, скачивается 20 лет.

@sweetmax797 что за сайт? впервые слышу

@usersupermengg348 Да, с VPN скачивать файлы , это настоящий кошмар. Этот сайт , китайская версия Huggingface, большинство моделей сначала выпускается там, а затем публикуется на HuggingFace

@usersupermengg348 Try Zapret, you don't have to use a VPN

@Eschelon I don't think that HuggingFace Hub got banned in Russia, more like Cloudflare Storage has been banned there and HF Hub uses it 😤

@usersupermengg348 Modelscope was made by Alibaba. It's a Chinese site, it's not very good because a phone number is required to access limited repos and use spaces

Это инфа для всех.

В zapret добавляешь в list-general.txt домены civitai (если надо) и hugging face (домен: huggingface.co). У меня после этого всё скачивается как прежде (У меня тоже Ростелеком). Вообще в целом если случается подобная штука, то первым делом кидаем все зависимые домены в zapret. Также если вы не знаете нужные домены сайтов или те, на которые ссылается сайт, то часто их можно найти через просмотр кода элемента. Таким образом у меня list в zapret увеличился в 2 раза т.к это часто помогает.

@sweetmax797 напрочь все кроме яндекса, вк и макса. пользуем свой впн через облако яндекса и workplane который больше не работает на яндексе, но позволяет кинуть вирт порт на любой внешний впн и так получаем трафик. был бы гугл доступен, там на мезе крутится впн с 2007г. у нас полностью закрыли доступ через всех провайдров. приезжие немного офигевают от нашего чебурнета. я умудряюсь сделать по 2ПБ месячного трафика ростелекому и 150Тб мегафону. работают в паре + когда есть линк на скайлинк сканирую его. аналоговое оборудование осталось еще с 2003г, работает исправно в любых сетях с любыми ключами и любыми кодеками. пол города кормлю инетом во зло властям! скоро соберу свой радиохаб для вафли и если не посадят, в городе будет независимая вафля. nokia 7730 наше все, спасибо связям в майках, всегда помогают с оборудованием!

@Eschelon Это просто эпик, как тебя ФСБ ещё не прихлопнуло, само по себе двойной эпик, особенно во время войны, когда у них все глаза на стреме. Будь осторожен, братан, я лично от малейшего тормоза в инете с катушек слетаю, так что полностью в теме с этой хренью и желанием найти выход. Для таких как я ты реальный герой. Держи ухо востро и не рискуй зря

Update comfyui will appear among the templates

already included in png images, just open a demo png image in your comfyui or use the old flux1 dev workflow it works

Эта модель только что выпущена и говорит более точно на русском. https://civitai.com/images/111679055

No, it's a new model trained for scratch

512x288 2x upscale = 2048x1152 takes 45 seconds

x4 to 4096x2304 (when model is loaded already) took 38 seconds.

Hi, you mean it is not a complete FP8 version of mistral small?

@sweetmax797 I don't think either bf16 or fp8 is a complete version!

@sunweixi1993786 flux has dtype of bf16 https://huggingface.co/black-forest-labs/FLUX.2-dev/blob/main/text_encoder/config.json
now, stop playing teacher and tell us what you mean by "not a complete version" ? do you mean that it is not a direct quantized model of Mistral Small, or what?

@sweetmax797 Yes, what I mean is that both of these are quantified versions. The files provided by black forest labs are ten files from 0001-0010 to 0010-0010. If these shard files (ten files) are completely merged together, the capacity is definitely around 40G. It can't be 35G of the "small" bf16

@sunweixi1993786 oh my God :) ok. thank you for the news and good luck!

Hello my comrade from Russia! You can change the link to the hf-mirror.com mirror but before doing so create a system environment variable named HF_ENDPOINT and set its value to https://hf-mirror.com. Good luck in your creative endeavors!