HighCWu

Amateurish AI male player, focus on face swap, anime colorization, voice conversion and novel continue writing.

[HighCWu @HighCWu]

I wonder if vibe coding is helping accelerate this direction.

Building a native compiler is far more feasible today than it was a few years ago, so we're starting to see more experiments in this space.

Curious how others see it.

[HighCWu @HighCWu]

1/

TypeScript seems to be heading in two very different directions.

Perry has been exploring native compilation. Vercel just released ScriptC. Meanwhile, the TypeScript team is focused on making TS-to-JS much faster.

These feel like two very different futures for TypeScript.

[HighCWu @HighCWu]

3/

That never became part of the TypeScript roadmap.

Instead, the team has focused on making TS-to-JS faster, most recently by rewriting the compiler in Go.

To me, these are two fundamentally different directions.

[HighCWu @HighCWu]

@nearcyan I still remember that time when BigGAN was at its peak. It took me a lot of time to convert the weights of Shawwn's anime model trained in Tensorflow to PyTorch. Now Tensorflow is a thing of the past, while TPUs seem to have become a crucial cornerstone for Google's AI training.

[HighCWu @HighCWu]

I would like to know how nano banana can perform style transfer, just like gpt4o's style ghibli. So far my attempts to perform this style transfer have not been successful🤔

[HighCWu @HighCWu]

@sebastienrama @ostrisai I have plans but not now, because it is much larger than the flux model, and even if it is quantized, it is difficult to train on my 16GB graphics card. I plan to observe the progress of the community.

[HighCWu @HighCWu]

I modify some code of @ostrisai /ai-toolkit to support training Kontext LoRA on 16GB VRAM.
Config: https://t.co/Kz3840BnhB
Example Config for training Jojo style: https://t.co/mywbTviAug
Link to GH: https://t.co/AItHl4aL6r

[HighCWu @HighCWu]

@bdsqlsz This looks like a quick HTML element modification. 😳

[HighCWu @HighCWu]

@sebastienrama @ostrisai I'm glad that helped you.

[HighCWu @HighCWu]

@Teslanaut @ostrisai I found that offloading text_encoder_2 cuts VRAM by ~2.6GB. Without it, training uses ~15.1–15.5GB (incl. display). With `unload_text_encoder` config parameter in ai-toolkit, usage can drop to nearly 12GB.
You can try it yourself, but it's best to use a system without a GUI.

[HighCWu @HighCWu]

@Teslanaut @ostrisai maybe it's possible with text encoder offload or prompt embed pre-cache. I'll try the idea later

[HighCWu @HighCWu]

@blxcknxir @ostrisai Good question. It relies on 4-bit quantization to train on 16GB, versus the usual 8-bit on 24GB. The quality should be very comparable, but I can't do a direct A/B test myself. Would love to see if someone in the community can run a comparison!

[HighCWu @HighCWu]

@okaris @ostrisai I've PR before I post this😉

[HighCWu @HighCWu]

@Jinbin_Bai @BlinkDL_AI As an independent researcher, I do not have sufficient funding to attempt large-scale model training. At most, I can only conduct some preliminary experiments to test ideas.😪

[HighCWu @HighCWu]

🚨 What if you removed linear layers from Transformer blocks?
Surprise: It still works. 🤯
I replaced linear layers with embedding layers — inspired by @BlinkDL_AI 's DeepEmbed.

💡Introducing Embformer, embedding-weight-only Transformer.
👉 0.1B Demo: https://t.co/OVLEu4xpRc

[Kohya Tech @kohya_tech]

FramePack、動画全体ではなく過去フレームから次のフレームを推論していく動画生成モデル。13Bモデルで、6GB VRAMで60秒、1800frameの動画を推論できる。過去フレームのpack方法を工夫してコンテキスト長を固定して動画の長さによらずO(1)。Project Pageがわかりやすい。
https://t.co/n2WVjvlGQN

[HighCWu @HighCWu]

gpt4o's image generation are very popular. I think if it was just style transfer, into studio ghibili style or others, I've done it a long time ago. If combined with PCM to accelerate the model, I can generate many images of different styles with the same content in one second.

[HighCWu @HighCWu]

I am very happy that my idea of #controllora has been applied in #sdxl's #controlnet (although the person who implements and trains it is not me). I've been playing with ip adapter recently, and I feel that it has great prospects in style transfer. (from ffhq to ghibli)

[HighCWu @HighCWu]

Since the previous style transfer, my experiments have made some new progress. Now, combined with ip-adapter and control-lora-3, I can input reference images and convert my images to the style of the reference images without adjusting the strength like img2img.

[HighCWu @HighCWu]

I have released 3 versions of control-lora so far. They are all based on the image diffusion model. Due to insufficient resources, I have not trained https://t.co/IkYLfoRLZl for video. I am glad that someone used the same idea as mine in the video model . Long live open source🎉

[toyxyz @toyxyz3]

https://t.co/T6RDaMF05h