v4.0
這次刻意加入了更多真實的皮膚紋理與材質細節，少了過度磨皮的濾鏡感，或許有些習慣精緻網美風的人會覺得畫面不夠討喜。不過對我來說，保留這些微小的瑕疵與真實感，才能讓光影更有層次，也更貼近現實的質感。
By deliberately introducing more realistic skin textures and raw material details, the overly smoothed "beauty filter" look is toned down. Some who prefer polished, glamorous aesthetics might find it less instantly appealing. However, I believe preserving these subtle imperfections adds greater depth to the lighting and brings the rendering closer to tangible reality.
今回はあえて肌のキメや質感など、よりリアルなディテールを強調したため、フィルターをかけたような陶器肌に慣れている方には、少し好みが分かれる（刺さりにくい）描写かもしれません。ただ個人的には、こうしたわずかな粗さや自然な質感を残すことで陰影に深みが生まれ、より現実に近い生々しい質感が表現できると考えています。
피부 결이나 사물 질감 등 사실적인 묘사를 한층 강조하면서 과도한 보정 필터 느낌을 덜어내었기에, 화사하고 매끄러운 미형 연출을 선호하시는 분들께는 다소 호불호가 갈릴 수 있습니다. 하지만 이러한 미세한 디테일과 자연스러운 질감을 살려야 조명의 깊이감이 더해지고, 현실에 더 가까운 무게감을 줄 수 있다고 생각합니다.
v3.3
重新訓練的 Lora，使用了全新資料集，並加入 BS4 的新 Lora 權重，混合原版模型之之後，在與 v2.0 分權重融合。整體的多樣性會略微提升，並消除了 v2.0 部分材質表現不佳的狀況。
The retrained LoRA utilizes a brand-new dataset and incorporates the new BS4 LoRA weights; after being blended with the base model, it is merged with v2.0 using weighted fusion. This results in slightly improved overall diversity and resolves the issues with poor texture rendering found in v2.0.
再学習されたLoRAは、全く新しいデータセットを使用し、新しいBS4 LoRAの重みを取り入れています。このLoRAはベースモデルとブレンドされた後、加重融合（weighted fusion）によってv2.0とマージされます。その結果、全体的な多様性がわずかに向上し、v2.0で見られたテクスチャ描画の不備も解消されています。
재학습된 LoRA는 완전히 새로운 데이터셋을 활용하고 새로운 BS4 LoRA 가중치를 적용합니다. 이 LoRA는 베이스 모델과 결합된 후, 가중치 융합(weighted fusion) 방식을 통해 v2.0과 병합됩니다. 그 결과 전반적인 다양성이 소폭 향상되었으며, v2.0에서 나타났던 저조한 텍스처 렌더링 문제도 해결되었습니다.
v2.0
根據 v1.0 再加入新的臉部資料，並加入 2000 張細節與動態表現的資料。
New facial data has been added based on v1.0, along with 2,000 images of detailed and dynamic expressions.
バージョン1.0をベースに、新たな顔データが追加され、詳細かつダイナミックな表情を捉えた2,000枚の画像も追加されました。
버전 1.0을 기반으로 새로운 얼굴 데이터가 추가되었으며, 세밀하고 역동적인 표정을 담은 2,000장의 이미지가 포함되었습니다.
v1.0
這是從 [Krea2 Turbo] AsianMix Lora 合併原始模型權重，並加入全新訓練的 TQD 以及全新資料集的 Lora，與原始 Lora 表現雷同但增加一點多樣性。
This is a Lora model that merges the original model weights from [Krea2 Turbo] AsianMix Lora, and adds a newly trained TQD and a new dataset. It performs similarly to the original Lora but adds a bit more diversity.
これは、[Krea2 Turbo] AsianMix Lora のオリジナルモデルの重みを統合し、新たに学習したTQDと新しいデータセットを追加したLoraモデルです。オリジナルのLoraと同様の性能を発揮しますが、多様性が若干向上しています。
이것은 [Krea2 Turbo] AsianMix Lora 의 원래 모델의 가중치를 통합하고 새롭게 학습한 TQD와 새로운 데이터 세트를 추가한 Lora 모델입니다. 오리지날의 Lora와 같은 성능을 발휘합니다만, 다양성이 약간 향상하고 있습니다.
Description
FAQ
Comments (21)
This is a very interesting model. Is there a version that supports INT4?
Nope, just int8 convrot, nvp4 and gguf K4_K.
@hinablue That is a shame. Are there no plans to add INT4 to the schedule?
@amegyu Why you need int4?
@hinablue The Krea2 models generally require a very large amount of VRAM. My setup uses an NVIDIA GeForce 30-series card with 12GB of VRAM, so int4 models are preferable. I would really appreciate it if you could consider releasing an int4 version as well.
@amegyu OK, give me some time :P
@hinablue Thank you, I'm very happy!
@amegyu Finally, int4 is done.
Wow, the INT4 quality is absolutely stunning, too! Thank you so much! It makes me so happy.
Thank you, I really appreciate it. I do want to mention that, although I'm sure it makes a lot of people happy, this checkpoint seems to have a "preference" for large/balloon sized breasts. Even requesting 'small breasts' or 'petite' doesn't seem to work consistently. I'm just trying to get "realistic" looking... doesn't even have to be "small" hehe
Haha, maybe next version 🤣
Use a breast size slider. There's a lot from the LoraHolic guy.
i tried many famous model here, especially those famous already in ZIT era,.... asian utopian isn't great in ZIT era, but this Krea 2 v2 is really nice .... nicely balanced .. thanks for your effort.
个人来说相当满意这个模型的胸部表现，特别是在不完全裸露的情况下，线条相当符合物理规律，很好的体现了该有的柔软和重力感。
Is it possible to download bf16 of your wonderful model somewhere? Unfortunately I can't use quantized versions on my setup yet.