AI Painting Parameters: What Actually Matters After Two Years of Tinkering

I'll be honest: when I first started with Stable Diffusion, I thought parameters were an afterthought. Write a good prompt, hit generate, done. The result was a lot of random luck—a hundred images, maybe three usable ones, and no idea why.

It took a long time to gradually understand what the key parameters actually do. This article isn't trying to be a comprehensive parameter guide. I just want to explain the ones I've genuinely figured out through real use. Articles that list 15 parameters with two paragraphs each aren't useful—you won't remember them, and even if you do, you won't know when to adjust what.

So I'm trying a different approach: imagine you're sitting next to me while I adjust parameters, and I tell you what I'm thinking.

Seed: The Only Parameter Worth Taking Seriously

Seed is the only certainty you have in AI painting.

Other parameters control "how to draw." Seed determines "what gets drawn." The same prompt with different Seeds produces different compositions, different poses, different lighting directions.

My workflow looks like this:

Exploration phase: Seed goes random. Just blast out a batch and see how it feels. This phase isn't about details—it's just checking whether the prompt direction is right.

The moment you get a composition you like, lock the Seed immediately. Then polish on top of it—tweak the prompt, add detail descriptions, swap style words. Because composition, once lost, might not come back with the same prompt.

After locking the Seed, change only one variable. Today I only tweak the character's expression. Tomorrow I only change the background. If you change three things at once and something goes wrong, you have no idea which one caused it.

For a while I didn't know Seed could be fixed. Every time I got a good image I'd just screenshot it, never able to reproduce it. It's like having a beautiful dream you can't remember when you wake up.

CFG Scale: Higher Is Not Better—That's the Biggest Misunderstanding

CFG controls how "obedient" the AI is. The higher the CFG, the more strictly the AI follows your prompt.

Sounds all good, right? I used to think so too.

Then I discovered that cranking CFG too high makes images explode—not broken-artifact explode, but colors oversaturated to the point of eye strain, contrast cranked to unnatural levels, like someone slapped an Instagram filter on your photo and maxed everything. Highlights on faces can blow out to pure white with no detail.

Too low, and the AI starts doing its own thing. You ask for blue sky and white clouds, and it gives you an abstract painting. The composition might be interesting, but you have no idea what it's supposed to be.

I use 7 most of the time. For anime-style images I might push it to 8 for that vibrant, punchy look. For softer, more photorealistic scenes I'll drop to 6 and let the AI breathe a little.

The takeaway: 6 to 8 is the comfort zone. Stay there for most cases. If you see people online running CFG at 12 or 15, I won't judge, but if your images look overexposed or the colors are muddy, try lowering CFG first.

Sampler: Obsessing Over This Is a Waste of Time

Sampler selection is probably one of the biggest newbie obsessions. Euler a, DPM++ 2M Karras, DDIM, UniPC… the names alone are annoying enough to memorize.

My experience: pick one that works and forget about this setting.

I use DPM++ 2M Karras. The reason is simple—it's not bad, not slow, not error-prone. Like a family sedan. Nothing exciting, but it gets you anywhere.

I sometimes swap to Euler a for fun. It has bigger step-to-step variation, so the images have more "surprise"—good for the exploration phase when you're blasting out batches to feel things out. But I wouldn't trust it for final output because the results are less predictable.

People spend hours testing how different samplers affect the same image. My take: the better your prompt, the smaller the sampler difference becomes. A well-written prompt won't produce terrible results with any sampler. Conversely, a garbage prompt won't be saved by any sampler.

If you must have advice: newbies, use DPM++ 2M Karras, set steps to 28-35, don't touch anything else, go practice your prompts first.

Resolution: Don't Get Greedy

This is the deepest hole I've ever fallen into.

Early on, I always thought bigger was right. Straight to 1024×1024 or higher. The result? Single-character images were fine. But put two or three people in the frame and it's over—two heads on one body, or three legs.

The reason isn't complicated. Models are trained at 512×512 (SD 1.5) or 1024×1024 (SDXL). When you significantly exceed that size, the model gets "confused"—it doesn't know how to fill the extra space, so it repeats, distorts, and fabricates.

The correct workflow: render at base resolution, then upscale afterward.

Specifically, for SD 1.5 models, render at 512×768 or 768×512. If you need it bigger, use Hires Fix to scale 1.5x-2x. The resulting large image has way better detail and structure than rendering at high resolution directly.

SDXL models are natively 1024, so scaling up a bit from there is fine.

For character close-ups I usually use 512×768 (portrait), run Hires Fix to push to 768×1152—very stable. For wide scenes I use 768×512, same approach.

Denoising Strength: The Soul of img2img

If you do a lot of img2img, then Denoising Strength is the most important parameter after Seed. Period.

It controls "how much to change." 0 means don't change. 1 means completely redraw. Simple in concept, subtle in practice.

My common scenarios:

Minor tweaks, keep the overall image: 0.25-0.35. Like fixing a messed-up face, or adjusting some colors while keeping the composition intact.

Change style but keep composition: 0.45-0.55. This range is interesting—the composition stays, but the texture and brushstrokes completely change. Like the same sketch being repainted by a different artist.

Major changes: 0.65-0.75. At this point the image is very different from the original—only the rough outline and positional relationships remain. Good for "I have a composition I want to redraw."

Don't go above 0.8. Past 0.8, you might as well generate from scratch. Too little information is preserved, and often the result has nothing to do with the original.

Hires Fix denoising strength is a common trap. Many people run Hires Fix with the default 0.7 denoising, and the upscaled image ends up completely different from the original. For Hires Fix, denoising should be 0.35-0.45—it's for refining details, not redesigning.

On Negative Prompts

Negative prompts are, honestly, overhyped by a lot of people.

There are all these "universal negative templates" online with dozens or even hundreds of terms. I'm not denying they work, but if you paste a hundred negative words every time, you can get weird side effects—certain elements you actually wanted get suppressed too.

My approach: scene by scene, add selectively.

For character work, hands are my biggest concern, so bad hands, extra fingers, missing fingers always go in. For landscape work, hand issues don't exist, so those terms are unnecessary.

Some negative words affect image style too. If you add "3D render" as a negative term, the image gets forced toward a hand-drawn direction—which isn't always what you want.

My personal negative template is short. Seven or eight items, mostly anatomy-related problem terms. I add targeted ones for special cases.

Details Nobody Talks About

When batch generating, 4 images at a time is enough. Some people say generate 8 or 16 at once, but VRAM aside, 4 images is enough to tell whether the prompt direction is right. If it's right, lock the Seed and refine. If not, change the prompt and try again.

Clip Skip—if your model recommends setting it to 2, set it to 2. Don't touch it. Many anime-oriented models were trained with Clip Skip 2. Changing it makes the style wrong.

I basically never turn on Face Restoration. CodeFormer and GFGPAN sometimes produce fake-looking faces, like someone else's face got pasted on. I'd rather re-generate than use repair—unless the composition is great and only the face is broken, in which case I'll occasionally use it.

My Workflow Summary

This is my current standard process. It might not work for you, but it's efficient for me:

Quick prompt + random Seed + 20 steps → blast 4 images, check composition direction. Find one I like, lock the Seed, start adding detail descriptions and style words. Once composition is solid, bump steps to 30-35 for high quality. Need a bigger image? Add Hires Fix. If some local area isn't right, crop it out for img2img with low denoising (around 0.3) to fix details.

The core idea: composition first, then details, then upscale. One thing at a time.

Final Word

Parameters—read ten articles, still worse than adjusting them ten times yourself. The effect of many parameters depends on your model, your prompt, your subject matter. Someone else's "optimal settings" might not work for you.

My advice: change one parameter at a time, compare the results, understand what it does. Don't change three parameters simultaneously and then ask "why did it get worse"—you won't know which one to blame.

If you genuinely understand Seed, CFG, resolution, and denoising strength, you're already ahead of most people who just fill in parameters randomly. Everything else, learn it when you encounter it.