Han Yoon PRO

hi, i have tried Qwen-Image-Lightning-4steps-V1.0.safetensors in V100, the inference time cost 90s.
What surprised to me is that your Qwen-Image-Edit_Fast-Presets outcome the image just in 15s.
So, how did you speed up the inference time with original QWen model?
Very appreciate for your reply!!!

Without some additional context, its impossible for me to know the cause of the difference in inference speed.

• Are you hosting it on HF Spaces as a ZeroGPU powered instance?
• How are you loading the lightning LORA?
• Are you inferencing an external model for prompt enhancement, or what model are you using for captioning?
• And et cetera... Wish I could be of more help. Feel free to provide more info, or even better, a link to the repo/space.

To those that are completely confused by this post, the HF Space that is being referenced: https://huggingface.co/spaces/LPX55/Qwen-Image-Edit_Fast-Presets

Edit: privacy clarification -- while user prompts are printed in logs for debugging purposes, they are not saved nor stored, neither are the inputted + generated image(s). We do not read the logs unless there is active dev/debugging going on. When there is activity on our end, the logs get wiped clean every build / pushed commit so they are typically never read. Don't be weird.

This space is still a very much WIP, there are a lot more ideas and features planned but on the backburner. We have experimented with various other prompt types which have shown pretty impressive results, so stay tuned. Lastly, the prompt enhancement feature is being served by an outdated model for the time being as we look into a handful of potential lightweight SOTA models.

Been meaning to respond to this, but as a fellow proud neurodivergent, of course it took me more than 10 days. I have several ideas that I've been beginning to explore. Would love to hear your feedback -- if you're interested that is.

Awesome idea! Just a heads up - its case sensitive ;)

Cool to see HF contributions and GitHub contributions stacked on top of eachother -- interesting patterns and insights that would have otherwise been unnoticed.

Guess we'll just have to get R1 to re-write all those API endpoints and specs for us 😂 On a more serious note, totally agreed on OpenAI's dominance and influence in anything LLM-related APIs. Definitely not going to change overnight, considering practically 90% of AI dev is currently using their specs.