Joined December 2012
264 Photos and videos
Pinned Post
YES!
introducing Realtime Edit. now, like with Nano Banana, you can edit images with complex instructions in real time. join the beta 👇
105
784
13,752
1,579,017
vicc reposted
Just pushed support in AI Toolkit to train @krea_ai Krea2 with reference images. In my testing, it learns edit concepts just as fast as an actual edit model would. More in 🧵
16
28
310
19,962
loons 🎈 IMG: @krea_ai 2 -> Qwen Image Edit -> @thesystms INFL8 LoRA MOV: @Alibaba_Wan 2.2 I2V T2V Vace -> @thesystms melt/transition/INFL8 LoRAs TimeSlice MP3: @suno EXE: @ComfyUI
7
19
215
5,679
Jungle Vision™ @krea_ai 🦍 @ComfyUI 🦍 @AdobeAE 🦍 @thesystms 🦍 @Alibaba_Wan 🦍 @suno
3
6
68
2,060
vicc reposted
2
13
638
vicc reposted
Training image models requires a surprising amount of Tinkering: prototyping reward models, training a prompt expander, and creating the RL environment. We love the creativity from @krea_ai!
our technical report is out. deep dive on the data, architecture, and training techniques used to create Krea 2. krea.ai/blog/krea-2-technica…
6
10
179
40,627
vicc reposted
thanks to the Thinking Machines team, we used Tinker to prototype our reward models and train the prompt expander via RL. for more information, read the full technical report on the data, architecture, and training behind Krea 2 👇
Training image models requires a surprising amount of Tinkering: prototyping reward models, training a prompt expander, and creating the RL environment. We love the creativity from @krea_ai!
4
5
128
15,631
My first go at training a style LoRA on @krea_ai 2 with @ostrisai AI Toolkit and hoooooly shit it's a good model to train! Animated in @ComfyUI with Wan 2.2 I2V T2V Vace with a melting LoRA I've been working on 💦
11
7
153
4,178
vicc reposted
Excited to see that @krea_ai shipped UniFusion design with Krea 2! More frontier models now recognize that multimodal semantics are never encoded in just a single layer. In the paper we showed that not only is Layerwise Attention Pooling better than all other conditioning methods, VLM hidden layers are perfectly capable of *reconstructing* images, thereby enabling editing without VAE inputs whatsoever. Glad to see UniFusion proven effective at scale. Congrats to @viccpoes and team, it’s a great model 🚀
today, we release the open weights of Krea 2. welcome Krea 2 Raw and Krea 2 Turbo, an undistilled model from mid-training meant to be fine-tuned, and a fast distilled version with a wide aesthetic diversity. read the details below 👇
1
2
18
16,663
vicc reposted
Spend a lot of time writing this over the past week. If you want juicy details on our training stack and infra. Go read it! krea.ai/blog/krea-2-technica…
9
28
163
20,506
vicc reposted
Krea 2 [officially] landed on the hub and its really good🔥 2 checkpoints a LoRA collection 1) Pretrained Raw: ideal for fine-tuning 2) Distilled Turbo: ideal for inference (w/ loras too!), fast & high quality Coming with day-0 diffusers support LoRA training and inference 🧨 👇 huggingface.co/krea/Krea-2-R… huggingface.co/krea/Krea-2-T…
8
76
6,950
vicc reposted
Krea 2 trained a style I couldn't get into any other model. Midjourney included. It's about time an open source model got there. I've been sitting on this dataset forever. It's of full-page plates from an early-1900s book of watercolor illustrations (via Public Domain Review). The texture and faded tonal gradients with soft bleeding edges tripped up every model I tried... none could get the aesthetic right with decent prompt adherence. Krea 2 nailed it.
30
33
634
67,044
vicc reposted
Ostris AI Toolkit has day 0 support for training LoRAs on Krea 2. The model has beautiful aesthetics and finetunes very well. 🧵
today, we release the open weights of Krea 2. welcome Krea 2 Raw and Krea 2 Turbo, an undistilled model from mid-training meant to be fine-tuned, and a fast distilled version with a wide aesthetic diversity. read the details below 👇
17
25
235
21,378
vicc reposted
Musubi TunerでKrea 2をday 0サポートしました。H2D only block swapとの組み合わせでほぼ速度低下なく12GB VRAMから学習可能です。またトークン順の最適化、flash attentionのGQAネイティブ対応などで高速化を図りました。 ドキュメントはこちら:github.com/kohya-ss/musubi-t…
2
30
145
20,633
vicc reposted
I've had this idea for a technique in my head for a while so I thought I'd give it a shot - the goal is to take depth-based slices of different pieces footage and merge them together somehow Could be a fun one to explore some more
12
9
120
4,412
vicc reposted
I made a training adapter for @krea_ai Krea2 Turbo that will allow you to train a LoRA on the turbo model directly without breaking down the distillation. This works the same way as my Z-Image turbo adapter.
16
22
280
14,696
vicc reposted
krea 2 turbo
6
1
30
821
btw we're hiring
today, we release the open weights of Krea 2. welcome Krea 2 Raw and Krea 2 Turbo, an undistilled model from mid-training meant to be fine-tuned, and a fast distilled version with a wide aesthetic diversity. read the details below 👇
7
3
97
9,843
vicc reposted
today, we release the open weights of Krea 2. welcome Krea 2 Raw and Krea 2 Turbo, an undistilled model from mid-training meant to be fine-tuned, and a fast distilled version with a wide aesthetic diversity. read the details below 👇
93
247
2,099
742,813
vicc reposted
if you want to help train SoTA text to image models you should apply to krea btw
today, we release the open weights of Krea 2. welcome Krea 2 Raw and Krea 2 Turbo, an undistilled model from mid-training meant to be fine-tuned, and a fast distilled version with a wide aesthetic diversity. read the details below 👇
7
7
87
9,324
must read
Spend a lot of time writing this over the past week. If you want juicy details on our training stack and infra. Go read it! krea.ai/blog/krea-2-technica…
19
4,125