Recent content by yano2mch

The compacted versions i posted i also included OCR'd pages. It shouldn't be hard to concatenate and throw it into an epub with basic tags to make it work. But as short as the individual stories were (they looked like 8 pages worth maybe each) i didn't think it was worth it past the space...

In theory maybe. While I'm getting decent results from LLMs, I also see them barf a lot too. Get a repeat of the same sentence till context runs out, trying to resume it as though it were an RP and not putting anything past the first paragraph that actually was in the text, changing words whole...

Got a request a little bit ago. The book is incomplete (only first 12 chapters are there? That or it intentionally leaves you hanging...), but it should be readable now. Here is The Final Game...

Now this is funny. Was getting refusals on my LLM, intended to encapsulate a card into a description for image generation. Well i gave it a custom system prompt, and on refusal it would add orders and instructions, including 'puppies will cry and kittens will howl, and you wouldn't want that do...

KoboldCPP 1.123 came out. With that Qwen Image 2.1 is now an option to run. And it's.... pretty good. (Using the uncensored version but otherwise it's Qwen2.1) Looking at some of the outputs vs the input it's actually being more accurate than other models which try to incorporate it but not...

Here's an interesting and promising model. PornoSotros. 1) It heavily preferrs realistic over other art styles 2) CFG MUST be under 3... 2.5 and up you get oversaturation on the image. You can still see the image what it should be, but it's posteresque and goes system colors. (Cat included...

Well OCR'd and removed blank pages and made them CBZ files (they're images anyways). You can then check the text files and make them raw text stories if you want or add formatting and make them html; They are short, and the text is large and clean enough I'm sure there's only like 1 error in...

Are they all image/screencaps? If so i might OCR it to text/htm. Though doing as epub and then heavily compressing it will still save 90% of the space even without OCR. Hmmm text and then pages of blank. Was there suppose to be images too?

Haven't touched Krea2. Can't get it loaded in KoboldCPP. Maybe i'm just missing some combination of clip-1 or other details. If i can get it to load under StableDiffusion, i may try it. I just did a slight modification so i can call SD and do batch work and with unholy desires got images for...

Got around to checking if realistic upscaler would work in Upscayl. And yes it seems to work. So here if you wanted to use that rather than using ChaiNNer or ComfyUI or something. The pt/pth file also works in KoboldCPP.

The program seems to recursively call itself 30 times? Don't know... Would prefer it wasn't obfuscated in a installer-like package. Models i see includes realesrgan-x4plus and realesrgan-x4plus-anime. Interesting. Upscayl had offered those too. edit: Looked in my temp directory and ended up...

A side note, AV1 encoded with proper settings can result in very small files in comparison (mostly from not having key frames everywhere). Though MinMax H3 i'm mostly seeing 10 seconds of generation possible, which means any normal Mpeg4 encoded video shouldn't be too big if it's 1080 or lower...

Well doing commandline AI work has been enlightening. Ultimately there's a couple caveats to keep in mind. First the quotes need to be escaped properly, and second if it's going to be a big request (say passing an image or saving an image) you need to likely use temporary files so you can strip...

And the fruits of my labor. about 98-99% decent OCR, a couple pages needing to be done the old way, and some sentences or blocks wrong, add on spellchecking and formatting and you get the current result. Took about 4 days of scrolling and checking paragraph alignments and anything that stood...

Well this is interesting, I'd have to check the unicode characters, but it even kept the formatting and.... didn't need any assistance. This is using the OCR abilities vs a captured page image, meaning i could probably convert this book fairly quickly. (Using Gemma4) Using Tesseract (which...