dalle-flow VS meadowrun

Jina AI is really into generative AI. It started out with DALL·E Flow, swiftly followed by DiscoArt. And then…🦗🦗*🦗🦗. At least for a while…

you can host your own https://github.com/jina-ai/dalle-flow

For example for all the multimodal stuff like clipseg and upscalers, I'm using isolated executors through jina flow: https://github.com/jina-ai/dalle-flow/tree/main/executors

clipseg is an image segmentation method used to find a mask for an image from a prompt. I implemented it as an executor for dalle-flow and added it to my bot yasd-discord-bot.

Merged into [dalle-flow](https://github.com/jina-ai/dalle-flow/pull/112) this morning and works on my Discord bot [yasd-discord-bot](https://github.com/AmericanPresidentJimmyCarter/yasd-discord-bot).

https://github.com/jina-ai/dalle-flow with ports of some code from https://github.com/lstein/stable-diffusion plus some stuff specific to my uses (mostly more exposed settings and meta data on the outputs).

dalle flow - a more advanced version of dall-e mini, running dall-e mega and a diffusion model (free colab), free

Another option is https://github.com/jina-ai/dalle-flow combines DALL-E Mini with some other image processing models, and they have a pre-built Docker image that you could run locally. However, because it loads additional image processing models, you'll need about 21 GB of GPU RAM which is more than a 2080 TI has. You could always try to edit their Dockerfile and re-build it to remove the other models.

For the second half of this article, we’ll use meadowdata/meadowrun-dallemini-demo which contains a notebook for running multiple models as sequential batch jobs to generate images using Meadowrun. The combination of models is inspired by jina-ai/dalle-flow.

If you’re anything like us, though, you’ll feel compelled to poke around the code and run the model yourself. We’ll do that in this article using Meadowrun, an open-source library that makes it easy to run Python code in the cloud. For ML models in particular, we just added a feature for requesting GPU machines in a recent release. We’ll also feed the images generated by DALL·E Mini into additional image processing models (GLID-3-xl and SwinIR) to improve the quality of our generated images. Along the way we’ll deal with the speedbumps that come up when running open-source ML models on EC2.

So it is more reasonable to cache the download locally for up to 4 hours. That saves us 5–10 seconds on every run.