[EN]
An experimental high-rank LoRA (Rank 1280) for the FLUX.2 Dev architecture (~8B parameters). The model was trained on a curated dataset of 400 high-quality works of classical marine and easel painting. Its primary goal is to recreate the fundamental physics of 19th-century oil painting: dimensional glazing, the translucency of sea waves, soft atmospheric sfumato, and dramatic use of contrast.
Generation and CFG settings: CFG 0.0 (Recommended for the classical style): Reveals the pure physics of the LoRA's latent space. It yields maximum internal luminosity, soft misty haze, and signature wave transparency without shadow clipping. CFG 3.5: Strongly anchors the generation to the prompt and renders fine geometric details (useful for complex machinery and architecture); however, the brushstrokes become denser and more matte.
Prompt prefix: For optimal text encoder input, always start your prompt with the combination: oil painting, [your prompt]. Model weight: Standard default 1.0. Publication and renders: All visual materials in the release gallery are published in a 100% untouched state (Raw Outputs) straight from the generation pipeline, without any post-processing, retouching, or upscaling.
Recommended Settings: Parameter
Value Base Model FLUX.2 Dev (~8B)
LoRA Weight 1.0
CFG Range 0.0
(Classic/Glazing) 3.5
(Detailing)
Sampler: dpmpp_3m_sde_gpu
Scheduler: linear_quadratic
Steps | 30 – 60
Prompt Prefix oil painting,
[UK]
Експериментальна LoRA високого рангу (Rank 1280) для архітектури FLUX.2 Dev (бл. 8 млрд параметрів). Модель була навчена на ретельно відібраному наборі даних із 400 високоякісних творів класичного морського та станкового живопису. Її головна мета відтворити фундаментальні фізичні властивості олійного живопису XIX століття: багатошарові лесування, напівпрозорість морських хвиль, м’яке атмосферне сфумато та ефектне використання контрасту.
Налаштування генерації та CFG: CFG 0.0 (рекомендовано для класичного стилю) розкриває чисту фізику латентного простору LoRA; забезпечує максимальне внутрішнє сяйво, м’який серпанок і характерну прозорість хвиль без втрати деталей у тінях. CFG 3.5 міцно прив’язує генерацію до текстового запиту та відтворює дрібні геометричні деталі (корисно для складних механізмів та архітектури); однак мазки пензля стають щільнішими й більш матовими.
Префікс запиту: для оптимальної роботи текстового енкодера завжди починайте запит із комбінації: oil painting, [ваш запит]. Вага моделі: стандартна (за замовчуванням) 1.0. Публікація та рендери: усі візуальні матеріали в галереї релізу опубліковано в незміненому вигляді (Raw Outputs) безпосередньо з конвеєра генерації, без жодної постобробки, ретуші чи апскейлінгу.
Рекомендовані налаштування: Параметр
Значення | Базова модель: FLUX.2 Dev (бл. 8 млрд)
Вага LoRA: 1.0
Діапазон CFG: 0.0
(Класичний стиль / Лесування) | 3.5
(Деталізація)
Семплер: dpmpp_3m_sde_gpu
Планувальник (Scheduler): linear_quadratic
Кількість кроків (Steps): 30 – 60
Префікс запиту: oil painting,
[RU]
Экспериментальная LoRA высокого ранга (Rank 1280) для архитектуры FLUX.2 Dev (~8 млрд параметров). Модель обучена на специально отобранном наборе данных, включающем 400 высококачественных произведений классической маринистики и станковой живописи. Ее основная цель воссоздать фундаментальные физические свойства масляной живописи XIX века: многослойные лессировки, полупрозрачность морских волн, мягкое атмосферное сфумато и эффектную работу с контрастом.
Настройки генерации и CFG: CFG 0.0 (рекомендуется для классического стиля) раскрывает чистую физику латентного пространства LoRA; обеспечивает максимальную внутреннюю светимость, мягкую дымку и характерную прозрачность волн без потери деталей в тенях. CFG 3.5 жестко привязывает генерацию к промпту и прорисовывает мелкие геометрические детали (полезно для сложных механизмов и архитектуры), однако мазки становятся более плотными и матовыми.
Префикс промпта: для оптимальной работы текстового энкодера всегда начинайте промпт с фразы: oil painting, [ваш промпт]. Вес модели: стандартный (по умолчанию) 1.0. Публикация и рендеры: все визуальные материалы в галерее релиза представлены в исходном виде (Raw Outputs) напрямую из конвейера генерации, без какой-либо постобработки, ретуши или апскейлинга.
Рекомендуемые настройки: Параметр
Значение | Базовая модель: FLUX.2 Dev (~8 млрд)
Вес LoRA: 1.0
Диапазон CFG: 0.0
(Классический стиль / Лессировка) | 3.5
(Детализация)
Сэмплер: dpmpp_3m_sde_gpu
Планировщик: linear_quadratic
Шаги: 30 – 60
Префикс промпта: oil painting,
Description
Comments (13)
Is this a LoRA or a Checkpoint, it's Very large for a LoRA?
It is a LoRA, but an experimental extreme high-rank configuration (Rank 1280).
Because of the massive 1280 rank on FLUX.2 Dev (~8B parameters), the weight file is several GBs — much larger than standard rank 32/128 LoRAs. It captures deep stroke brushwork, physics, and layering without quantization compromises.
Use it as a standard LoRA at weight 1.0!
Pro-tip: If you find the LoRA too heavy to load dynamically, just merge it directly into your base FLUX.2 Dev model using ComfyUI or SD-WebUI! You'll get a standalone baked model with zero inference overhead and no need to carry separate LoRA weights.
@ORAKUL_STUDIO Sooooo you're giving me the usage right to merge with the main checkpoint? The Usage Right for Merge is marked as "No Merging allowed" :)
@SencneS Haha, that was just a default Civitai checkbox left enabled by mistake! 😅
You have my full official permission to merge it directly into base FLUX.2 Dev. For a massive 1280 rank, baking it into the checkpoint is actually the best way to run it with zero overhead. Go ahead and merge away! I'll update the page permissions as well.
@SencneS By the way, if you are engaged in training yourself or are just interested in accelerating the pipeline, visit me on GitHub: https://github.com/OrakulStudio/AI-Toolkit-Windows11
We here just the other day made an impossible breakthrough with this 1280th rank for FLUX.2 Dev on a single RTX 4090:
• Speed increased from the standard 150+ seconds to reactive ~39 seconds per step!
• Lowered the transformer offload to 0.65 (completely zeroed out PCIe bus congestion) and configured a manual reset of the latent cache.
• Zero OOM when saving and only 136 W consumption on FP8.
All the open code of tulkita and descriptions of optimizations are laid out there. Use it if you want to get the most out of iron!
@ORAKUL_STUDIO wow that is a big speed up, but I only have 16gb Vram, (4070ti super).
But I'll give you a thumbs up on there. I would say though that I do have some good ideas, just not a coder. Anything I code I vibe code. Test it, if something fails, I vibe code those fixes as well, along with ideas and concepts that have helped A LOT with my Workflow.
I created about 15 nodes that are highly customized for Krea2 or Ideogram 4. I was consider trying to release it for some Buzz, but CivitAI don't really have the options to do that.
btw, I vibe coded a LoRA optimizer that will use SVD to bring the size of LoRAs down while maintaining 99.95% of the data (or any value I enter) While running this LoRA was pumping out an average weight of 159, with a peak of 299, and a low of 131 Rank. I didn't finish it, because it's going to take about an 90 minutes, so I'll run it over night and see.
I had an idea a while ago, and had been fiddling a lot, then stopped the project, but this LoRA made me want to try it again, and it appears to work. Just didn't let it finish, I will tonight.
@SencneS That SVD compression idea sounds awesome! SVD pruning high rank matrices down to their core dynamic weights while keeping 99.95% variance is actually a super smart approach, especially for fitting monster models onto 16GB cards like the 4070 Ti Super.
Let me know how the overnight run turned out! Really curious to see the quality drop if any vs the size reduction after your script processes the Rank 1280 weights. Keep vibe coding!
@ORAKUL_STUDIO
Processed 160 LoRA layers.
Average Layer Rank: 6144.0 → 1092.2
Processed at 99.95%
Original Size: 14880.06 MB → New Size: 12818.67 MB (13.9% space saved)
Processed 160 LoRA layers.
Average Layer Rank: 6144.0 → 662.5
Processed at 99.50%
Original Size: 14880.06 MB → New Size: 7846.27 MB (47.3% space saved)
Processed 160 LoRA layers.
Average Layer Rank: 6144.0 → 493.5
Processed at 99.00%
Original Size: 14880.06 MB → New Size: 5835.37 MB (60.8% space saved)
The interesting thing about the 99% is it highlights layers that are just not really used.
New Ranks List (160 layers):
[299, 76, 363, 735, 145, 53, 71, 168, 331, 176, 561, 798, 227, 99, 112, 198, 417, 373, 679, 820, 334, 150, 197, 325, 557, 415, 740, 839, 236, 262, 261, 374, 528, 520, 721, 795, 249, 291, 281, 323, 467, 543, 680, 752, 263, 270, 340, 383, 417, 519, 628, 679, 240, 286, 351, 361, 360, 489, 582, 598, 223, 286, 303, 198, 617, 675, 592, 665, 390, 301, 534, 694, 519, 699, 562, 734, 563, 731, 594, 724, 585, 692, 541, 682, 569, 709, 581, 687, 567, 418, 566, 679, 593, 692, 626, 750, 636, 742, 650, 755, 622, 756, 621, 713, 604, 715, 567, 693, 556, 688, 396, 509, 561, 721, 518, 685, 467, 624, 391, 634, 358, 615, 379, 660, 401, 699, 431, 681, 429, 717, 415, 629, 435, 572, 392, 660, 370, 668, 418, 689, 432, 585, 491, 569, 454, 505, 454, 280, 103, 88, 456, 572, 445, 542, 427, 516, 428, 501, 375, 454]
There are some layers here that are under Rank 100, and some that are way upwards of 750.
The difference between 100% and 99.95% only saved a couple of GB. But I'd be interested in your analysis of the 99.5% - If there is zero change or super tiny changes at the 99.5% that is a significant drop in size. While the 99% is even smaller for half a %, the second half of that % doesn't save much compared to the first half.
Either way - I will not upload these to CivitAI out of respect for you. But if you'd like I can throw them up on HugginFace and send you a private link.
@SencneS Hey man, huge thanks for this incredible analysis! The layer-by-layer breakdown [299, 76, 363...] is absolutely fascinating seeing how FLUX dynamically allocates rank weight across layers proves just how much capacity a high-rank model can hold.
Honestly, we hadn't gotten around to compressing our builds yet due to lack of time, but your research and enthusiasm totally infected us! Inspired by your post, we're rolling out two big releases:
Tonight: The Rank 1280 (200 steps) model we'll publish both the full-size uncompressed version and an officially SVD-compressed variant.
Late tonight Tomorrow morning: The absolute monster Rank 1536 (150 steps), which is literally baking right now on our RTX 4090! It will also come in both full and compressed versions.
Please definitely send over the private HuggingFace link for your test builds I’d love to test them out and check the generation quality. As for public releases, we'll upload our official SVD versions under our main profile, but I'll make sure to credit you for the inspiration and analytical breakdown!
Also, a quick offer: Since you're running on an RTX 4070, I can set up a private HuggingFace repository specifically for you, where I'll upload all our experimental models pre-compressed and tailored for 12GB VRAM hardware. That way, you'll get early access to everything we train, and we can swap private links there. Let me know if you're down!
@SencneS Actually, if you'd like to publish your SVD version on CivitAI yourself, go for it! Just mention Orakul Studio as the original creator and link back to us. Appreciate your respect and great work!🤍
@ORAKUL_STUDIO Published here - https://civitai.red/models/2945204/flux2-dev-aivazovsky-xxi-svd?modelVersionId=3334887
I put all them in a single post with variants.
@SencneS Hey
Thanks again for setting up the SVD release so cleanly and with so much respect for the source work.
Following up on our earlier chat I'm going to set up a private GitHub repository for SencneS and send you an invite. I'll be uploading all my full, uncompressed high rank master models and raw experimental weights directly there so you can grab them first hand for testing and SVD optimization.
Quick heads-up on what's cooking in our pipeline:
Rank 1536 is practically ready and dropping in a couple of days.
Rank 2048 is next in line for testing we're going to stress test the stack to see where the absolute limit of the hardware and precision forcing lies.
I'll DM you the repo access link as soon as it's initialized. Let's keep pushing the limits!