Architects' tools

My first LORA for Flux. Will share it soon, probably next week. Just need to check and choose between all the models that were made with the same dataset. I already have my favorite, but it's better to test twice. Thanks Ismail Seleit for recommending Colab, because locally it's hard to train something for Flux now

that's it for now, but we'll be back to Flux this week.

One of the advantages of Flux is the processing of complex detailed prompts; the model can handle them quite well, although not without problems. Nevertheless, in terms of prompts, there is a lot of room for tests. There is, among other things, the possibility to write several prompts and mix them. In this example, the foreground with the field and animals and the background with the museum building are described separately. Two consecutive samplers are used in workflow, as the first one does not always show the desired result (the second one, by the way, not always either, sometimes you need to choose the right mixing factor). The two images in the post show what happens after the first and second sampler.

ComfyUI Essentials node package has Flux Sampler Parameter, which is like KSampler, but with the ability to set several values for any parameter through commas, thus simulating the result of XY Plot script. For example, this way you can find out that guidance around 1.5 - 2 returns the most realistic results, and the optimal combination of sampler+scheduler, as it seemed to me, is euler+beta. but you can test it yourself.

The same with the same prompt for Dev model. Here the result is already, at least in my opinion, noticeably better. Including the design of bridges in most cases is more logical, it seems as if you can even walk on them, without stumbling into dead ends, and this is great, because we are still talking about basic models.

Let's start with the basic workflow for Schnell model. There are several ways to make it. First of all, you can download a ready-made one from comfyui developers' github , where you can also find links to the Flux models (there are also links to lightweight versions of the models at the bottom of the page, if your graphic card can't handle the full ones). But I'm somehow used to the standard ksampler, so here is a variant of the workflow through it. The images, as you can see, not so good, but this is still a basic model, without any LORA and ControlNet, which we will later layer on this workflow.

I guess it's time to talk in detail about Flux, after all, the model has been around for almost a month and we can start to try it in #comfyui let's start from the base. there are three models in this Flux family: 1. Schnell (i.e. fast from German) works like SDXL Turbo, it renders the image in a small number of steps (4-5), but the quality suffers. This model is free and you can use it as you wish. 2. Dev. Good output image quality, but you need 20 to 50 steps. This model is free to use for non-commercial purposes. 3. Pro. Can not be used for free and can not be downloaded, you can only try online, but it is not particularly interesting, so for now we will not pay attention to this model, despite the fact that in terms of quality it is the best of the three.

plugin for #comfyui that adds a sidebar with a library of nodes and some other handy things to the programme interface: https://github.com/Comfy-Org/ComfyUI_frontend

I've made a page on civitai, if it's more convenient to download from there: https://civitai.com/models/627685/geometric-minimalism

file of the model itself and workflow for tests:

I'm training my next LORA model, it's still raw, but it's not a given that I'll keep working on it, so I'll post it as is. the idea is to have it generate intricate but geometrically clean building shapes. #comfyui

I tested Flux 1 on Friday. everywhere people write that it's a kind of Midjourney killer, but objectively, no, it's not, and it doesn't even replace SDXL yet. at the same time, Flux is much better than any other open source model in its pure form, i.e. on the simplest workflow, when you don't have ControlNet, IP-Adapter, Lora (all this is not available for Flux yet), but only a regular text prompt. All in all, the potential is good, and if Lora will be well trained on top of it, it will be great, but for now it's just interesting, no more than that. on the downside: the model is heavy and to work with it you need a graphic card with 16 gb of memory. works only in #comfyui , the other interfaces, as usual, later. download here: https://comfyanonymous.github.io/ComfyUI_examples/flux/

last project for today designed by Nasim Nouri #education

another project made by Angeline Susanto, Instagram : @eugenia.angie #education

Today I will show you images from the presentations of the workshop participants, which we (SA lab and I) conducted as part of the annual DigitalFUTURES event. first work from Ksenia Gorina (nickname @as_found in instagram). recordings of the classes are here. #education