FasciumKROMA
FasciumKROMA is a premium dual-variant checkpoint family, meticulously merged from the blazing-fast Kroma v0.2 (Krea 2 Turbo) architecture. By integrating FASCIUM's proprietary aesthetic tuning directly into the full model weights, this release delivers exceptional generation quality at unprecedented Turbo speeds.
This release is divided into two distinct branches—Realistic and Art—allowing you to choose the perfect tool for your specific creative workflow without sacrificing the speed and efficiency of the underlying Turbo merge.
📸 Variant 1: FasciumKROMA Realistic
This branch is engineered for hyper-realistic generation with a strong focus on the "Real Women" aesthetic. It excels at rendering authentic, diverse human portraits with natural skin textures, subtle imperfections, and classical photographic lighting.
Aesthetic: Classical portrait photography and painting vibes, soft cinematic lighting, and elegant compositions.
Key Features: Highly detailed skin and fabric textures, natural diversity in age and body types, and a strict adherence to clean, elegant beauty (no tattoos). It captures subtle suggestions of form without crossing into vulgarity.
Example Prompt:
mid-body portrait of a 35-year-old woman, natural skin texture with subtle freckles, soft cinematic lighting, classical painting aesthetic, elegant pose, highly detailed, photorealistic, 8k resolution, natural hair, no tattoos, detailed environmental elements, sharp focus, shot on 85mm lens. Negative prompt: tattoos, explicit nudity, plastic skin, oversaturated, deformed, ugly, bad anatomy, text, watermark, thoughts, expressions.
🎨 Variant 2: FasciumKROMA Art
This branch is tailored for deep stylization, digital illustrations, and high-end concept art. It takes the structural understanding of the base model and pushes it into vibrant, stylized, and highly expressive artistic territories.
Aesthetic: Digital painting, concept art, rich color palettes, and dynamic, expressive compositions.
Key Features: Distinct brushstrokes or digital rendering styles, masterful use of color theory, and the ability to generate highly stylized characters, environments, and fantasy/sci-fi concepts with incredible artistic cohesion.
Example Prompt:
masterpiece digital painting, a mystical forest guardian, intricate armor with glowing runes, dynamic pose, rich vibrant color palette, dramatic chiaroscuro lighting, highly detailed concept art, expressive brushstrokes, artstation trends, stylized illustration, 8k resolution. Negative prompt: photo, photorealistic, 3d render, plain, low quality, bad anatomy, deformed, messy, text, watermark.
Description
FAQ
Comments (11)
Огромная благодарность за ваши труды! Стоит ли ждать от вас int8 моделек?
Точно нет. Начал разбираться и не вижу смысла - fp8 мало чем уступает но конверсия убьет много времени и сил
@FASCIUM на моей 3090 разница большая, 00:18<00:00, 2.25s/it против 00:12<00:00, 1.53s/it
@Grufrin У меня тоже 3090, но есть нюанс. Я работаю только в Forge, там всё не так просто с int8, а поскольку я квантизую только в чистую через скрипт python есть нюансы. разобраться можно, но на всё времени просто не хватит
@FASCIUM я могу посмотреть скрипт и попробовать разобраться, насколько сложно туда добавить INT8 и что именно придётся менять со стороны Forge
@Grufrin Я знаю что нужно, но это долго и нудно. Просто лишняя работа, а у меня другие задачи
Hello guys! If I understood your problem with int8 correctly, then do it yourself - here is the tool - https://github.com/qskousen/ggufy?tab=readme-ov-file if you can’t build it yourself, here is the releases page - https://github.com/qskousen/ggufy/releases
The most sensible option is to download the bf16 version and quantize it however and into whatever format you like. I won't release anything I can't test myself.
This is why I always download only BF16, and only then transfer it to the format most convenient for me.
@FASCIUM not everyone has enough VRAM and RAM to quantize Krea-based models. IMHO 4-bit quants have to be provided by default.
Here are the parameters of the system on which I am running BF16, yes it is slow, but it works, the W4A 4 format works the fastest.
OS: Manjaro Linux x86_64
Kernel: 7.2.3-2-MANJARO
CPU: AMD Ryzen 7 5800H with Radeon Graphics (16) @ 4.463GHz
GPU: NVIDIA GeForce RTX 3060 Mobile / Max-Q 6GB
Memory: 27147MiB / 31483MiB

















