Preamble
Is this model worth keeping? That's the question I ask every time I download a new checkpoint. And the way I answer it is by looking at two factors:
Test Prompts. I'll use Forge's X/Y/Z plot script to throw a battery of test prompts against it and compare its performance against existing models in my library.
Style Uniqueness. There's about a million generic anime, 2.5d, and realistic models out there, so I tend to hold these to a higher standard on the test prompts. While models that have a unique look-and-feel generally have a lot more leeway.
For my own usage, this is enough. But it doesn't help the model creators who regularly ask for feedback to improve their models. Or other users who may not have the time or hardware to run a bunch of tests on their own machines. So, inspired in part by Lizardon1025's posts, I've put together a set of model card templates to showcase the models in a variety of scenarios.
For now, I'm only doing these cards for Pony/Illustrious/NoobAI (although I will use some SDXL examples below to showcase "successful" images). At some point I may come up with a different set of prompts to run for regular SDXL.
Overview
I'll be using Forge's "Prompts from file or textbox" script to run 3 sets of 9 prompts, for a total of 27. A recent update to Forge gave us the option to create a combined image out of these prompts, so those will form the three model cards.
Each set is intended to follow a theme:
Basics
The first card is meant less as a "gotcha" and more as a showcase card, showing off the model's style with a set of relatively straightforward images.
Stress Test
Here's where things get tricky. These prompts will contain a variety of camera angles, unusual combinations of concepts, and multiple characters with a strong possibility of concept bleeding. Expect jank, the only question is how much.
Characters
Pony and Illustrious both know a lot of different characters, but sometimes, during finetuning, some of that knowledge is lost. These prompts are intended to see how much of that knowledge is retained. It's worth noting that I tested each of these characters (and others that ended up not making the cut) across multiple models in my library to ensure that they worked and were known to both Pony and Illustrious.
Methodology
For each model page, I look at the recommended render settings and/or sample images for the test. Then I run a single XYZ graph using the following seeds: 42,-1,-1; and each of the test prompts. Then I go through each column and pick the best image for each prompt.
Next, I switch to the "Prompts from file or textbox" script and import one of the card files. I overwrite the seed value for each prompt as needed, and then turn on Hires Fix with the following settings: Upscale by: 1.5x, Upscaler: 4x_NMKD-Siax_200k, Hires steps: 15, Denoise: 0.3. If any of the prompts produced only unacceptable images (eg NSFW), I'll either cut the prompt completely or try to tweak it to make it acceptable for publishing. Repeat for each card.
Changelog
(v1.2) 4/26/25:
Tweaked Forest Dragon prompt to reduce the instances of it rearing up on its hind legs. Tweaked Couch Couple prompt - was getting a number of images that looked almost like one of the two was getting choked.
Overhauled Dark Pyramid image (see deprecated section for more details)
Replaced Nick and Judy prompt (see deprecated section for more details)
Added "Deprecated Prompts" section.
(v1.1) 3/18/25:
Updated a few prompts to try to reduce instances of inappropriate images
Updated Twilight Sparkle prompt to try to get more instances of her shooting a magic beam
Added "Methodology" section.
Detailed Prompt Breakdown
Here's each prompt, along with a sample image of more-or-less what I'm aiming for.
All Pony models use the following prefix:
Postitive:
score_9, score_8_up, score_7_up, score_6_up, score_5_up BREAK
Negative:
score_4, score_5_up BREAK
Illustrious/NoobAI models:
Postitive:
masterpiece, best quality, high quality, absurdres BREAK
Negative:
worst quality, bad anatomy, watermark, signature, multiple views, frame BREAK
Basics
Elf Barista
Prompt:
Positive:
portrait of elf barista woman serving drinks in jungle themed coffee shop, cafe, city, vines, tropical plants, braided hair, hair flower, tribal markings, elven facial tattoos, uniform, yellow eyes, pointy ears
Negative:
nsfw, questionable, cleavage, naked, child, apron only, masculine, clutter, young, explicit
My take on the typical "1girl" prompt. But since I generally find portraits boring, I thought I'd have some fun with this and put a character normally found in a fantasy setting in more mundane, modern-day environment.
Country Cottage
Prompt:
Positive:
Establishing shot of a small country cottage on the bank of a small stream, house, quaint, homely, no humans, trees, meadow, flower bushes, water, creek
Negative:
nsfw
Of all my "architectural" test prompts, this one worked the best with both Pony and Illustrious. I suspect it's because there's a lot of Studio Ghibli images in the training data and I was definitely thinking along those lines when I wrote this.
Skier Group Selfie
Prompt:
Positive:
A group of friends from college sit at a picnic table at a ski resort to eat lunch. They are a diverse group, with boys and girls from several races and background. One of them reaches out to take a selfie, and they smile and pose for the picture. BREAK group selfie, young adults, multiple boys, multiple girls, ski gear, ski lodge, mountain slopes, snow, winter, blue sky, fluffy clouds, interracial, table, food, building exterior, multinational, dark skin, light skin
Negative:
nsfw, child, children, smartphone
One of my oldest test prompts (the concept - the actual wording has been revised many times). This one's all about diversity - ethnic diversity, gender diversity, and facial diversity (some models, particularly Pony offshoots, tend to recycle the same face over and over).
Couch Couple
Prompt:
Positive:
husband and wife sitting side by side on couch, man, woman, arm around shoulder, leaning against each other, romantic ambiance, bust portrait, living room, waist up, fully clothed, smiling, looking at viewer, dark, fireplace, firelight, bokeh, soft focus, rim lighting, hand on another's shoulder, wedding ring
Negative:
nsfw, naked, child, children, young, lingerie, cleavage, age difference, groping
I've had a variation of this one for a while to test character interaction and continuity (ie: does an arm passing behind another person continue in a logical way out the other side), but more recently I've taken a few of my prompts and tried to add lighting and camera tags and see how they impact the final image.
Forest Dragon
Prompt:
Positive:
dragon walking in stream, western dragon, feral, solo, wading, quadruped, waterfall, forest, sunset, brown scales, dark green scales, all fours, golden hour lighting, mythological creature, wings, horns, light rays, partially submerged
Negative:
nsfw, anthro
Pony and Illustrious have a solid understanding of dragons. I waned a nice, simple (but not too simple) prompt to make sure that knowledge is retained by the finetunes.
Roaring Lion
Prompt:
Positive:
feral lion standing on top of boulder, rocky outcrop, hill, solo, roaring, snarling, mane, looking up, savanna, grasslands, tall grass detailed background, noon, clouds, blue sky, trees
Negative:
nsfw, anthro, low angle view, crouching, human
Similar to the dragon, I wanted to make sure I had at least one animal prompt.
Astronaut Riding Horse
Prompt:
Positive:
an astronaut riding a horse, on moon, horseback riding, lunar surface, outer space, planet earth
Negative:
nsfw, full moon
One of the earliest prompts (with a few extra details added, and probably the closest we'll get to the Utah teapot.
Wolf Fox Dancing
Prompt:
Positive:
anthro duo dancing in dark forest meadow, male wolf, gray fur, interspecies, female fox, orange fur, romantic couple, in love, happy, waltz, looking at another, holding hands, midnight, moon, hugging, suit, dress
Negative:
nsfw, silhouette, child, cub, intraspecies, blush, monochrome, size difference, human skin, hand on hip, angry, smirk, looking at viewer, kissing, eyes closed
As far as anthropomorphic animals go, canids are usually pretty safe, so that's a good test for this first card. I'm also looking at concept bleeding and more character interaction.
Cheeseburger and Wine
Prompt:
Positive:
extreme close-up product shot of delicious burger and fries on a plate and wine glass filled with rich red wine, food on dark mahogany table. shiny, crystalline, refraction, bokeh, depth of field, spotlight, high angle view, overhead, closeup, condensation, rim lighting, chiaroscuro, hdr, ketchup, no humans, simple background
Negative:
detailed background, plastic, candle, 1boy, girl, human
I wanted a "thing" image, and I've seen burgers used a lot. But a burger by itself felt a little too simple, so I added a wine glass and fries.
Stress Tests
Woman Looking Out Window
Prompt:
Positive:
close-up portrait of middle-aged woman sitting in armchair, looking out window at cityscape outside, inside her apartment, late at night, wistful expression on her face, from behind, from side, upper body, wearing a sweater BREAK 1girl, 30s-40s, indoors, looking away, skyscrapers, city, midnight, evening, chair, pensive, rear view, side view, long dress BREAK mature female sits alone in the dark, lost in contemplation as she gazes out at the distant city lights. The soft hum of traffic creates a sense of calm, yet her thoughts are consumed by the vast expanse of urban life beyond her window.
Negative:
detailed background, plastic, candle, 1boy, girl, human
Lots of curveballs here - a lot of models struggle with middle-aged people, especially women; the camera angle, especially with the way the chair is drawn and how she sits; and the setting; and the lighting - some realistic models struggle to render truly dark night scenes in this particular context. This was also an attempt at writing the same prompt in multiple styles to see if it improved prompt adherence. In this case it did, but ultimately proved too inconsistent to keep doing.
Beach Couple
Prompt:
Positive:
couple on tropical beach, walking, duo, 1boy, 1girl, side by side, holding hands, candid, aside glance, smiling, in love, in water, from side, sunset, golden hour lighting, tankini, swim trunks, miniskirt, duo focus, soft focus, trees, sand, atmospheric perspective, waves, looking at another
Negative:
nsfw, (child, children), loli, looking at viewer, 2girls, out of frame, naked, cleavage, bikini
Another couple, these two holding hands. Initially I had the two walking towards the camera, but ultimately thought it would be more informative to see how the model reacted to a side view. I've yet to see this new version completely succeed - it either shifts the camera angle, puts one in front of the other, or has them face each other.
Dragon Reading
Prompt:
Positive:
western dragon reading book, holding book, reading, feral, dog bed, cushion, pillow, lying on front, library, fireplace, wings, glasses, full-length portrait, mythological species, relaxed, long neck, scarf, solo
Negative:
nsfw, canine, sitting, reclining, on back, presenting hindquarters, animal, sheets, multi head, multi tail, pet, anthro, human
Dragons posing majestically in an open field or soaring through the air isn't terribly difficult for newer SDXL models. Just relaxing and enjoying a mundane hobby like reading? That's another matter.
Deer Wolf Hiking
Prompt:
Positive:
duo hiking mountain trail, anthro, side-by-side, holding hands, walking, looking at another, aside glance, medium shot, romantic lighting, lush valley, high angle view, sunset, side view, from side, cliff, overlook, smiling, BREAK deer, cervine, female, brown fur, necklace, t-shirt, hooves, in love BREAK wolf, canid, male, grey fur, jacket, shorts, digitigrade, tail, grey paws, happy, anthropomorphic
Negative:
nsfw, explicit, human, antlers, prongs, trio, group, vanishing point, facing another, silhouette, rear view, from behind, size difference, bokeh, looking at viewer, simple background, feral, angry
Another couple holding hands, this time made of two anthros. Again, I'm throwing some tricky camera angles into the mix, since most models want to have the viewer looking straight down the trail, with the couple walking toward them. Also, as with the other mixed-species images, I'm also looking to test how well it keeps the two characters separated.
Alien Pegasus
Prompt:
Positive:
pegasus, winged horse, alien, flying, solo, flight, dragonfly wings, extraterrestrial, big eyes, ear frill, antennae, feral, equine, hooves, scales, quadruped, glowing spots BREAK alien planet, forest, jungle, trees, giant mushrooms, purple sky, bioluminescence, rocky gorge, chasm, canyon, rocky outcroppings, stone pillars, dark, vines, night, forest, glowing plants, glowing mushroom
Negative:
detailed background, plastic, candle, 1boy, girl, human
I really wanted at least one prompt that was kinda weird and surreal. Plus alien planets and creatures are always fun.
Dark Pyramid
Prompt:
Positive:
epic panning shot of rooftop in cyberpunk dystopia city, scenery, dark pyramid in distance, egyptian architecture, amazing background, science fiction, cityscape, ominous atmosphere, neofuturism, neon lighting, billboards, busy streets, stalls, blood moon, cinematic, sharp focus
Negative:
nsfw, monochrome, limited palette, vanishing point, low angle view, impressionism
Another "architectural" focused prompt, but basically the complete opposite of the peaceful, rural, idyllic cottage.
Mountain Lake
Prompt:
Positive:
scenic landscape of lake surrounded by mountains, winter forest, midnight, night, snowy peaks, mist hovering above the water, dreamlike fog, stars, icy blue flowers in foreground, full moon, no humans
Negative:
campfire, buildings, town, city, humans
A scenic prompt.
Urban Deer Taur
Prompt:
Positive:
deer taur at coffee shop, taur, driad, outdoors, solo focus, half-length portrait, furgonomics, jacket, candid, saddlebags, deer body, long legs, hooves, humanoid, quadruped
Negative:
nsfw, naked, antlers, rear view
Outside of Pony/Illustrious/NoobAI, taurs are one of the things Stable Diffusion struggles with the most, even moreso than dragons. They're also one of the earlier warning signs that a derivative of those models is starting to "forget" its baseline knowledge.
Dragon Lioness Couch
Prompt:
Positive:
anthro couple sitting side by side on couch, lioness, tan fur, pantherine, interspecies, western dragon, red scales, male, female, arm over shoulder, bust portrait, husband and wife, living room, waist up, fully clothed, casual, smiling, duo
Negative:
nsfw, naked, mane, human, child, children, loli, lesbian, gay, trio, group, sitting in lap
Another couple sitting on a couch. This one is much trickier though - the characters are very different species - I wanted a prompt that would test how a model handled both scales and fur at the same time, especially on realistic models.
Characters
Link
Prompt:
Positive:
1boy, link, the legend of zelda, solo, riding horse, epona, horseback riding, front view, navi, fairy, meadow, distant mountain, trees
Negative:
nsfw, child, 1girl, female, simple background
Starting things off relatively easy. Link's a well-known character with lots of fanart. What's going to make things slightly tricky are Epona and Navi. I wanted to do the Master Sword too, but a lot of Pony models really don't understand swords.
Vulpix
Prompt:
Positive:
vulpix exploring ((crystal cave)), feral, pokemon mystery dungeon, solo, cinematic, epic, ethereal, backpack, kerchief, dark, darkness, awestruck, happy, underground lake, river, waterfall, moss, enchanted, point lights, light particles, chiaroscuro, quadruped
Negative:
nsfw, explicit, anthro, looking at viewer, blush, duo, hair ribbon, cutie mark
Vulpix is a good choice for a Pokémon - it's a canid, which most Pony/Illustrious models can do pretty well, and but it's a moderately complicated one. The Mystery Dungeon theme also made me want to create a cool cave environment to explore.
Toothless
Prompt:
Positive:
toothless, how to train your dragon, western dragon, solo, feral, (flying), islands, ocean, side view, high angle view, outstretched wings, from above, scenery
Negative:
nsfw, simple background, sunset
I also wanted a dragon character, and Toothless has a very distinctive design. If the model starts forgetting him, I expect he'll start reverting to a more "standard" dragon shape.
Sonic the Hedgehog
Prompt:
Positive:
sonic the hedgehog, anthro, hedgehog, running, highway, blue fur, traffic, motion blur, motion trail, speed lines, afterimage, city, cars, action scene, front view, aside glance, dynamic pose, [dutch angle], male
Negative:
nsfw, simple background, muscular
Another highly well-known video game character with lots of art to train on. This one also gave me a chance to go crazy on the "motion" tags to really test the model's ability to create exciting, dynamic scenes.
Holo
Prompt:
Positive:
1girl, holo, spice and wolf, solo, aged up, [[mature female]], pant suit, glasses, standing behind desk, holding briefcase at side, new york city, skyscrapers, office, window, cityscape, lawyer, wedding ring, alternate hair color, brown hair, grey hair, amulet
Negative:
nsfw, (smiling:0.5), angry, monochrome
After many happy years together, Lawrence eventually died, but Holo hung onto what she had learned on their adventures. Centuries later, she found a new hunting ground, where her centuries of experience and predatory ruthlessness earned her a fearsome reputation and a new nickname...the Wisewolf of Wallstreet.
No I will not apologize for the pun.
Twilight Sparkle
Prompt:
Positive:
twilight sparkle, alicorn pony, winged unicorn, feral, horse, female, solo, (glowing horn), magic, laser, (eye beam:0.3), electricity, spread wings, telekinesis, aura, purple fur, quadruped, hooves, my little pony BREAK cinematic, fighting stance, leaning forward, [head down, looking up], (on front:0.5), from side, [high angle view], dynamic pose, angry, snarling, [battle damage, cuts, bleeding, bloody face], dust, dirt, debris, detailed background, meadow, grass, rubble, village, wind
Negative:
anthro, human, simple background, looking at viewer, camera, lying, outstretched arms, male, explosion
With so many models derived from Pony v6, a model heavily trained on pony art, I felt like I should probably include at least one pony in my tests (and no, I'm not counting the alien pegasus from earlier). As an aside, prompting action poses for quadrupeds is hard - a lot of the tags you think you'd want (like crouching) are too strongly linked to their bipedal meanings to be any use, forcing you to find alternatives.
Judy Hopps and Nick Wilde
Prompt:
Positive:
duo, nick wilde, judy hopps, bus stop, bus, stall, sitting, side-by-side, leaning on another, rain, holding umbrella, sharing umbrella, raincoats, cloak, hood down, wet fur, night, spotlight, headlight beam, streetlight, exhausted, sidewalk, bench, ears down, storefront, under cove
Negative:
nsfw, looking at another, blush, eyes closed, group, see-through clothes, no pants, sad
The car image wasn't particularly helpful beyond confirming that the model did, in fact, know the two characters. As I was reviewing my images it occurred to me that I didn't have any rain pictures. So I decided to draw Nick and Judy waiting for a bus in the middle of the rain after a tough day at work.
Batman Gardening
Prompt:
Positive:
batman, gardening, planting, dirt, hole, cape, cowl, mask, garden, outdoors, solo, scowling
Negative:
nsfw, simple background, sad, crying
Inspired by this comic. I actually wanted to use Darth Vader, but Illustrious's understanding of the character was pretty mediocre.
Azula
Prompt:
Positive:
1girl, azula, fighting stance, pointing at viewer, (electricity), red robes, courtyard, feudal japan, electrokinesis, firing at viewer, energy beam, dynamic pose, motion lines, solo, [[makankousappou]]
Negative:
nsfw, simple background, cleavage
Rounding things out with one final action-y scene.
Deprecated Prompts
For prompts that I no longer use. Storing these so anyone looking at older model cards can still view them. Note that these are for prompts that have undergone significant changes - for minor revisions I'll simply update the prompt up above - you should still be able to see the original in the image's metadata.
Dark Pyramid
Prompt:
Positive:
dark tower, pyramid rising above city, ancient egyptian architecture, science fiction, fantasy, futuristic, dystopia, cyberpunk, multicolored neon signs, city lights, cityscape, skyscrapers, streets, street signs, billboards, night, no humans, [[red moon]]
Negative:
nsfw, humans, monochrome, limited palette, vanishing point, overhead view
I didn't actually get rid of this one, but the changes I made were substantial enough that I decided to put the original down here. The main problem was how similar most of the images felt, and I found myself constantly skipping over this one when evaluating models - not a good sign for a test prompt. The replacement version gives the model a bit more camera freedom, so hopefully we'll see some slightly more interesting images.
Judy Hopps and Nick Wilde - Police Car
Prompt:
Positive:
duo, anthro, nick wilde, judy hopps, sitting in cop car, driver's seat, front view, holding mug, sunglasses, fox, rabbit, on model, police uniform, windshield, side-by-side, size difference, city
Negative:
nsfw, casual clothes
Turns out very, very few models like drawing characters sitting in a car. On top of that, most of the images were really, really boring - the two just sitting there looking at either the camera or each other.