I just accomplished 6 months of self teaching.
There are two free sites I know of that I like that do what I like . . . that's pornworks and perchance. they are pretty damn accurate and powerful considering you learn to properly prompt. that was my first step. Once I could get the online webapps to give me what I want, I felt satisfied on that insight. Not saying there isn't more to learn, there is tons, and I am just starting out so, I use basic well written sentences to describe the general idea in one or two sentences. Then list details and use keywords separated by comma, for example.
Then I wanted to move on to learn more about the tech so I got a $150 ebay used EVGA GTX 1080TI. works perfectly for small models while keeping my budget way down. I can't see the point in paying more than retail no matter how good a GPU may be or new of a generation it is. I used LUMO to assist, prompting the (what I thought were the good) questions I needed to learn what I was trying to do. I run 32gb ram and its all ddr4, super cheap. NVME for fast OS speeds but I am using extra drives for large Model storage / Lora storage backup.
Got the hardware now so I set out to learn about LLM's and what the right terms were that I need to learn, know, understand, in order to guide me into the "ecosystem." Condsider the scope, I am only learning about image generation at this point. So my questions are quite specific. But I actually did start with text based LLM's, I run LM studio locally, with a nice model that is great with math and science, historical records, etc, it can code and it VERY good at it. I am using the Qwen3.5 9B Uncensored HauhauCS Aggressive model if you wanna know. Its good enough to get started I think.
So with that and LUMO, I was able to learn about what are frameworks, workflows, nodes, comfy UI, locally vs non local, difference between text gen LLM and image gen LMs . . . once I had a basic idea, I watched some youtube vids for a month or so, just to see what others are doing, and tbh, this wasnt so helpful here cause most people making guides and such, those vids were too old or software changed and I would get errors if followed along, and more impotantly they were too advanced and didnt explain the core basic stupid simple concepts that I lacked.
So I kept with the text LLMs and installed comfy UI and loaded my first model from RED-CIVIT.AI and it was a SD 1.5 model, to keep things small and easy for my GPU and also the workflow. It isn't as overwhelming as I initially thought, but only because I stuck with the more basic maps aka workflows. I use only the needed nodes and have nothing fancy going on. I didnt have expectations either as to what it was I was getting as an ouput. I just kept working around bugs, and software, installation issues, learning the ropes, using LUMO mostly because it can access the web and is more up todate - able to get latest updates on software, patches, changes, etc.
Most of my time so far has in fact been addressing properly creating the environments on my computer and keeping it all organized. Once I got the most basic setup working, running, I was able to start learning about the nodes, bugs I would have copying someone elses workflow, - that was a nightmare if complex btw, so I kept using LUMO to get me through that, using only the nodes neccessary to generate in SD 1.5, I have like 5 - 8 nodes or something iirc, um, the one with the lora I think. Maybe 6 nodes for just text to img ? I can't look now, I am actually training a lora right this moment.
So moving on, Once I got SD 1.5 down, I moved on to learning more about the different models, and what LORAs are, and styles or themes. Using Civit.AI, its quite easy to see what you get and they organize it well, so I stuck there for that. I found out what models and loras work for my GPU, I started downloading a few models for the few different architypes that exist, I have pony, krea2, SD 1.5, SDXL, Illustrious, so on and on, for each of those I got a few loras each, just a few, keeping directories clean and organized in a working VENV - this is for python in linux,
So then I learned and saved the most basic workflows for each model type - with / without lora, with / without img input along side text. So now I can use any of those models with and without loras to generate an image from text. Not to exclude learning about different ways to control the output, like UNET, text encoders, vae, and custom nodes.
I wanted to then train a character lora, so I set on about that, using LUMO to guide me again because it can access the net and has up todate information, I installed the training framework kohya_ss and got all that working, organized directory tree, etc . . . and am processing my V2 character lora this morning. I used for the first time this morning, my own lora I trained all night last night. Took 10 hours. Not bad I thought. So I asked again, how can I improve this after testing and using that lora. It works spot on but its powerful, too powerful in fact and has limited flexibility.
Making some adjustment to the initial training command, I am now running a training session to produce a more flexible lora but trained on the same dataset.
That is something I didnt mention, but I had to learn to create and process a dataset before training the lora . . . its just part of the process. There are of couse endless ways you can do this, but I chose a simple approach teaching me the core concepts, I am also using CLI not GUI so I get the extra experience with the terminal. It is more fluid and controllable, fine grain adjustment but a lil more complex to initially setup. You can use windows and kohya_ss but its a GUI and I am NOT using windows, like haven't been in over 10 years. Not going back now.
That said, I am very pleased with my progress as I do not listen to others bad advice when it comes to using AI as a guide, I ask folks and they all seem to have rather negative opinions about AI and I can see why, I also don't care so much for using Industry AI farms as they use lots of electricty, I like to use my local where I can. Just don't give up, use a tool, learn it well, and keep adding tools to your toolbox. Take everything someone says into consideration and don't be prejudice because you may lose out on something you could benefit from if you just hadn't looked away,
You will find what works for you and that is my advice.
edited for grammar or spelling etc
