i love this model

Reached 100 followers today Yay!!! Thank you to the ones following me 🫶💐 Onto 200 next🔥🚀

interesting read on parallelism some takeaways: - expert parallelism is the same as data parallelism for non-moe layers - for dp: the input batches of data are divided among the GPUs - for model parallelism: the inputs batches are replicated in each core.

would anybody be interested if I write a blog on How to optimize inference speed for diffusion transformer models like Wan?

I'm quite sad that Papers with code sunsetted. It was one of my most favourite platforms for papers

Sparse attention + DMD

another interesting blog from Owl they are really on fire🔥

A study on RL based AR image editing model Been searching for more RL based papers in multimodal gen space

Qwen nailed it at following complicated and long prompts.

Been talking about it for years. Wrote two books about it. The Holodecks are arriving! Huge.

Qwen's text generation capability is also quite good

here comes the next frontier in AI

realtime everyone

More demos of Genie 3 😮

jaw dropping honestly

Open source project similar to Genie 3. will be diving deep into this paper for sure

Fun workflow I was playing with last night 1) Kontext to remove Thor from the shot 2) Photopea to place Shrek 3) Kontext + Relight lora to blend him into the shot 4) Wan2.2 i2V to animate Very addicting... 😍 Prompts are written up in the corner. Wan 2.2 prompt below 👇

Trained a sidechain LoRA to compensate for the quantization precision loss when quantizing Qwen Image to 3 bit. It works well. This can be active during training and should allow us to fine tune Qwen Image on <24GB of VRAM. This can be done to all models.