Be good at something, and then use vibe coding to augment your talent, and automate it. I think that's really all it's good for from a product design point. It needs to be paired with something that isn't AI for it to work, and have a value proposition.
This is what I got
My go-to when I want to use a chatbot for info is for it to prompt some sort of roleplay and so far it has never refused to provide. Maybe it's a matter of how prominent the character is but I never hit a wall ever.
You could just prompt it to have the goal of defending a specific view point and you start poking holes. It's a good way to get your nogging jogging. I sometimes do it when trying to figure out things like design.
comfy is very good. I haven't played with the qwen model yet but it's highly-regarded. Flux in my opinion is too big and too slow to use. Illustrious is more or less only for anime
Looking forward to seeing you in the furry porn threads.
Qwen Image is good if you have the hardware for it. Flux 2 Dev is definitely too slow for general use, but Flux 2 Klein 4B and 9B are very solid models with both generation and edit capabilities that are a lot smaller and faster. I've been using 9B over Qwen Image for a while now.
x 18 x 4 x 1
It's KJNodes Ideogram 4 Prompt Builder, which is a bit different from regional prompting, in that the Ideogram 4 model is trained to understand JSON prompts containing bounding boxes. This node just lets you construct the JSON prompt visually.
In my experience it seems like each prompt affects execution time heavily regardless of whether it's regional or not. Each "conditioning" on the clip network nonlinearly increases it. The third adds more time than the second did, the fourth adds more time than the third, etc. Most other operations have a linear effect on execution time, like they just add a flat amount.
he doesn't understand teleoperation is impossible and that every factory that wants to run an AGI is going to need a server closet full of hardware to run the AI
I've done a bit of work in this area, and there's more than one way to skin this particular cat. Yes, you can use a VLA model to directly output motor signals and control a robot that way, but you can also use something like Code as Policies (which has progressed quite a lot since 2022, and the original paper now has nearly 2,000 citations). There's no real bottleneck for the latter, since the generated policy can run locally without any special hardware.
