I’m having trouble figuring out how to create images with the Novel AI Image Generator. I tried following the instructions on the website, but my results don’t look anything like the examples. Can anyone share tips or a step-by-step guide on how to get the best images? Any advice would really help.
Alright, straight up: NovelAI’s image generator is a fickle beast. It’s nowhere near as plug-and-play as the site makes it look, and their “magic” example pics are basically cherry-picked S-tier generations. If your images are coming out looking like melted crayons instead of cool waifus, join the club.
Big trick is that prompts matter WAY more than you’d expect. Descriptive and detailed prompt = better results. Instead of just writing ‘a cat,’ you gotta go ‘a fluffy tabby cat sitting on a rainy windowsill, soft lighting, photorealistic, intricate details.’ Also, add style cues – like “anime style” or “oil painting,” whatever suits.
Don’t ignore the advanced options. Play with the “strength,” “steps,” and “sampling” settings. Usually, cranking up steps gives you more coherent pics, but not always – sometimes you just get more gibberish. Low strength = more strict about your prompt, higher = it’ll go wild.
Tags: Experiment with positive tags (what you want) AND negative ones (what you don’t). “Missing fingers,” “distorted faces,” “blurry” etc, thrown into the negative prompt box, can save you a lot of mutated nightmares.
Batch generate like 8-10 at a time, then pick your favorite. MOST will be trash. It’s normal. Dump the garbage, keep the gems.
Lastly, don’t stress if your outputs don’t match their showcase. They’ve probably cherry-picked from thousands. Most of us are swimming in the uncanny valley.
TL;DR: Get specific, experiment, use neg tags, crank steps, and prep for a lot of trial and error. Anyone who says their first few were masterpieces is lying.
Okay, so here’s the thing: yes, @chasseurdetoiles is pretty much on-point about how picky NovelAI’s image gen can be, but honestly, I don’t buy that “prompts are absolutely everything” gospel. Yeah, the prompt matters, but I’ve had times where the most convoluted, paragraph-long prompt just gave me… more melted wax with anime eyes. Sometimes you gotta dial it way back and underspecify—just toss in a basic phrase like “dinosaur in victorian dress” and let the AI get creative, then nudge its failures with a couple retry attempts.
Also, the randomness factor is no joke. I know they say to batch generate, and I do, but sometimes the best image is from the first seed, other times it’s the seventh. What I found that helps is switching up the model or the art style you ask for (like swap from “anime” to “digital painting” and see what comes up). The “steps” and “sampling” settings everyone obsesses over? Meh. Sometimes cranking those up outright ruins a decent thing—too much clarity spoils the charm. Test both extremes!
Negative tags, though—total agreement there. “Extra limbs,” “deformed face,” “creepy eyes”—without those, you’re in nightmare city. But also, try tossing in “low quality,” “watermark” as negatives, sometimes it just starts spitting blobs otherwise.
One thing no one mentions: composition is weirdly important. If you want more than one character or something doing a specific action, break it into multiple phrases with commas instead of a wall of single concept tags. And if you really want to get fancy, try the “Weight” settings in advanced; that lets you emphasize specific elements by giving them numeric priority.
And don’t ignore the seed function! If you see something nearly good, re-roll with the same seed + minor prompt tweaks. Sometimes it lands.
So yeah—prompt is key, but randomness, tags, and tiny tweaks play a bigger part than they want you to think. Oh, and forget trying to match the “example” images—the gap is real. Trust the process, accept the failures, and meme the disasters. That’s half the fun, tbh.
Brutal honesty: most people using ’ seem to think “just crank out a prompt and BAM—anime waifu perfection.” That’s not how it actually goes. A lot of what’s missing from the advice so far is about planning, rather than just fiddling with tags and prompt word spaghetti.
So, get out of “shotgun mode” for a sec. Instead, try storyboarding what you want. If your image needs a specific pose, composition, or scene—draw it roughly yourself and use it as a reference if the ’ platform allows. Yes, it eats up time, but it’s the only way you’ll wrangle it into respecting your vision, especially for complex multi-character scenes.
Competitor-sided wisdom (from the earlier posters): definitely exploit negative tags (they help, but also sometimes the AI just ignores them). Batch gen? Meh. Occasionally, making eight messes and picking the least bad one feels like giving up, not improving. Sometimes, going slow—tweak, reroll, fix—delivers more consistently than the mass-generate method.
Pros for ': Ridiculously quick at making stylized, high-concept imagery. Great for exploring wild ideas or moodboards. When it works, it really works.
Cons: Unpredictability. Annoying habit of mangling anatomy or context. Takes tons of trial and error for specifics. Not reliable for client work without serious post-production cleanup.
Alternate idea: Try splitting up your prompt into modular components. Run the background first (“sunset over neon city, empty street”), save it, then prompt up your foreground character separately and composite them in Photoshop or GIMP after. Frankenstein the best bits together. Not the point-and-click holy grail you’re hoping for, but gets you closer to the examples you see flaunted everywhere.
In short—, like most AI art tools right now, is more “unruly artist” than “obedient paintbrush.” Build your vision in pieces, and plan for hacky clean-up.
And honestly, rival tools like those referenced by previous users each have their quirks; none are push-button magic, but ’ is decent if you’re prepared to wrangle, stitch, and sometimes just meme your mistakes.