/tech/ - Technology

test

Hello. Not sure if this is the right place to post this but baker is deleting at leat 50% of all posts. He is deleting Q posts, properly sauced relevant information, all new information, and any comments about deletions. I've seen bakers ignore or delete before, but never to this extent. Board is all but useless at this point. Morning bread up to noon. Hundreds of deletions and not talking about spam.

Thank you.

Part 2

547 deletions

supposedly a new baker now

we'll see

Thank you.

sir - how do I view the old 8ch.net/tech board?

I'm trying to post on this website from windows10 librewolf and im getting a 'server took too long to submit your post' error. im trying again from a windows10 edge and get same error. than i posted from winxp firefox and post went in fine. returned to win10 and get same error.

is there something in my win10 that could prevent my posts from going trough?

(this post is made in winxp)

Question + notes on AI image generation, since threads can't be created here.

Prompt vs reference, and why generators smear lettering:

1) Smeared lettering. Diffusion models don't spell, they draw letter-shaped texture. Past ~3 words the glyphs melt. Keep text short, quote it in the prompt, generate at the target aspect ratio instead of upscaling later, and use a text-render mode if the model has one. Otherwise make a clean plate and composite the type yourself - beats rerolling 40 times.

2) Oversharpen. The usual fix for a soft output is cranking sharpening or an upscaler at full strength, which gives halos on high-contrast edges and crunchy noise in flat wall/sky areas. Sharpen after resize, not before, mask it off skin and gradients, and stop while you still see grain rather than ringing.

3) Waxy skin. Too many denoise steps, CFG too high, plus "smooth/flawless/8k" tokens. Pores die first. Lower guidance, drop beauty adjectives, add slight grain, and for edits use low-strength img2img so original texture survives.

Rule of thumb: prompt for composition, reference image (img2img / ip-adapter style conditioning) for identity, manual compositing for anything that must be legible. Text alone will not hold a consistent character, product or logo across frames.

For quick side-by-side tests without a local install I've been using SoraLum, an AI image generator/editor covering image, prompt and video in the browser: https://soralum.com/image-ai/image-generator - useful for telling whether a failure is my prompt or the model.

Anyone here found a generator that renders more than a couple words of text reliably, or is compositing still the only real answer?