Zeno's Ziggurat


RPG characters with AI image creation

I claim no ownership or copyright of these images whatsoever. You may download and use them for whatever purpose you wish.


AI Basics: Combining Chat AI with GPT 2.5


I began experimenting recently with Chat AI as a sounding board for GPT 2.5 prompts. My prompting style is heavily influenced by the way DALL-E worked for the past few years, but DALL-E is no longer even part of my workflow (its not even available through Bing any longer). I wanted to see what the AI thought was the best way to prompt the image renderer, as well as see how well it could iterated and explain what was happening.


I started with an image from the halfling team I’ve been working on. Good archery shots are always prone to weirdness, so I started with Sal – the rogue archer – replicating a classic trick shot. I wrote a prompt the way I usually do, and asked Google AI how it would best be re-written as a GPT2.5 prompt:

What I got back was the below, together with a number of comments about what the AI thought did and didn’t work well about my prompt. Note that since I slapped the prompt together with cut-n-paste it had some issues – it really was asking for inconsistent results. But for purposes of this experiment its useful to see what the AI does with that.

PURPOSE:
A narrative fantasy character painting in the traditional Dragonlance art style, depicting a mischievous halfling archer in a tense but humorous scene, preserving a specific character design.

SUBJECTS AND ACTION:

  • Primary Subject: A 20-year-old male halfling archer. His facial features, hair, and exact likeness must be strictly preserved from the reference image @image1. He stands 3’6″ tall with a slightly stout build and classic hobbit proportions. He has long black hair, an easy smile, brown eyes, a stubble moustache, and is chewing a strand of wheat. He wears a classic brown Sicilian hat, a green shirt, brown pants, leather boots, and light leather armour.
  • Secondary Subject: A distant goblin standing nervously in the background, frozen in place with a bright red apple balanced precariously on top of its head.
  • Interacting Poses: The halfling is captured in a lazy slouch with his head cocked, sizing up the viewer with a mischievous look. He is carefully aiming a shortbow toward the distant goblin, though his relaxed posture shows complete confidence.

COMPOSITION:
A medium-wide environmental shot that establishes a clear foreground and background relationship. The halfling archer dominates the foreground, framed from the boots up. The nervous goblin is positioned clearly in the background, creating a humorous depth of field. The camera angle is straight-on, capturing the alignment of the bow arrow toward the distant target.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. Rich, earthy color palette dominated by forest greens, warm browns, and leather textures. Bright, natural outdoor sunlight filters through an overland fantasy setting, casting soft shadows that highlight the textures of his leather armor and the strand of wheat in his mouth.

CONSTRAINTS AND PRESERVATIONS:
Strictly preserve the exact facial structure, hair, and identity from @image1. Maintain distinct halfling/hobbit body proportions; do not make him look like a standard human or child. The shortbow must be actively drawn and aimed toward the background target. The apple must remain balanced on the goblin’s head. Avoid clean, modern digital art gradients; the texture should feel like a hand-painted canva

And the AI’s comments:

Key Improvements Made:

Style Details: Translated “Sicilian hat” and “stubble” into specific aesthetic guidelines that blend smoothly into the fantasy oil painting style.

Action vs. Pose Logic: Your original prompt mentioned him aiming and having a bow over his shoulder while slouching and looking at the viewer. This reformat clarifies that he is confidently aiming while maintaining that casual, cocked-head slouch.

Compositional Depth: Explicitly separated the halfling into the foreground and the goblin into the background so the AI creates a clear sense of distance for the apple-shooting joke.

The first thing I notice here is that it did a good job of catching the problem with the original – it immediately called out the inconsistent posing. The other thing that stands out is how organized and wordy it is. With DALL-E there was always a character count limit, and the AI explicitly lost track of things later in the prompt. So I would carefully word my prompts to avoid un-necessary filler words, and start with the most important elements and work my way down. In my original prompt I call out the style – to enforce consistency with other renders, then give a basic overview of the action, and then restate the particulars of the character’s description even though I’ve already provided an image reference.

Here the AI suggested organizing in strictly called-out sections – including the very interesting “CONSTRAINTS AND PRESERVATIONS” section. We shall see in a bit that it uses this section to do things like tell the renderer what NOT to do. And unlike DALL-E3 GPT2.5 actually appears to *listen* to “don’t do this” instructions! DALL-E listened like a five year old. If you said “don’t add pink elephants” the one thing you could be sure of is that it would add pink elephants. GPT2.5 – at least sometimes – actually appears to understand negative prompts!

So let’s see what that suggestion actually does:

Some definite issues here, as what he’s doing in this pic really doesn’t make a lot of sense. So instead of tweaking it myself, I just went back to the chat thread and said “That results in him looking at the camera rather than at the goblin, and not actually pointing his bow directly at the goblin.” The AI responds with the below – and two options for a better approach:

That happens because the prompt is giving the AI conflicting instructions by asking him to do two physical actions at once: look at the viewer and look at a distant target. When GPT Image 2.5 sees “sizing up the viewer” alongside “aiming his bow,” it usually prioritizes the camera gaze and compromises the archery posture.

To fix this, we need to choose one clear action. Here are the two best ways to reformat it depending on what you want the scene to feel like:

Option 1: The “Trick Shot” (Looking at the Camera)

Choose this if you want him to be a cocky show-off who is actively pointing his bow at the goblin without looking at it—aiming blindly while maintaining direct eye contact with the viewer.

Option 1: Blind Trick Shot Prompt

PURPOSE:
A narrative fantasy character painting in the traditional Dragonlance art style, depicting a cocky halfling archer pulling off a blind trick shot.

SUBJECTS AND ACTION:

  • Primary Subject: A 20-year-old male halfling archer. His exact facial features, hair, and likeness must be strictly preserved from the reference image @image1 . He stands 3’6″ tall, stout, with hobbit proportions, long black hair, stubble moustache, a brown Sicilian hat, green shirt, brown pants, and leather boots. He is chewing a strand of wheat.
  • Secondary Subject: A distant goblin standing nervously far in the background, frozen in fear with a bright red apple balanced on its head.
  • Interacting Poses: The halfling is facing the camera in a lazy slouch with a cocked head and an easy, mischievous smile. He is breaking the fourth wall, looking directly at the viewer. Crucially, his arms are extended outwards to his side, blindly pointing and aiming his drawn shortbow directly toward the distant goblin behind him without looking back at it.

COMPOSITION:
A medium-wide environmental shot with deep focus. The halfling archer dominates the foreground (boots up). The nervous goblin is perfectly lined up directly behind the arrow’s trajectory in the background.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. Rich, earthy color palette. Bright, natural outdoor sunlight casting soft shadows.

CONSTRAINTS AND PRESERVATIONS:
Strictly preserve the facial identity from @image1 . He must look directly at the viewer, but his bow and arrow must be fully drawn and pointing directly at the goblin in the background. No modern digital art styles.

Option 2: The “Focused Archer” (Looking at the Goblin)

Choose this if you want a classic, realistic archery shot where he is completely focused on hitting the apple, meaning his body is turned and his eyes are locked onto the goblin.

Option 2: Focused Archer Prompt

PURPOSE:
A narrative fantasy character painting in the traditional Dragonlance art style, depicting a halfling archer focusing intently on a target.

SUBJECTS AND ACTION:

  • Primary Subject: A 20-year-old male halfling archer. His exact facial features, hair, and likeness must be strictly preserved from the reference image @image1 . He stands 3’6″ tall, stout, with hobbit proportions, long black hair, stubble moustache, a brown Sicilian hat, green shirt, brown pants, and leather boots. He is chewing a strand of wheat.
  • Secondary Subject: A distant goblin standing nervously in the background, frozen in fear with a bright red apple balanced on its head.
  • Interacting Poses: The halfling’s body is turned in profile/three-quarter view toward the background. He is in a relaxed but steady stance, with his head turned sharply toward the goblin. His eyes are locked onto the apple. He is actively drawing his shortbow, pointing the arrow directly at the distant goblin’s head. A mischievous, confident smirk remains on his face.

COMPOSITION:
An over-the-shoulder or side-profile medium-wide shot. The halfling archer is in the foreground, with the camera looking down the line of his drawn arrow directly toward the nervous goblin standing in the background.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. Rich, earthy color palette. Bright, natural outdoor sunlight casting soft shadows.

CONSTRAINTS AND PRESERVATIONS:
Strictly preserve the facial identity from @image1 . He must NOT look at the camera; his gaze and the bow’s arrow must be pointing directly at the goblin. No modern digital art styles.

Option 2 there is more like what I wanted. And note the Constraints section – “He must NOT look at the camera”. So I feed that back to GPT2.5 and we get this:

That’s better, but he’s still not actually looking where he’s aiming. So lets ask the AI about that again: “Better. only problem is he’s still not actually looking at the goblin.”

If he is still refusing to look at the target in Option 2, it is because the image generator is struggling to override the original pose from your reference image, which shows him smirking directly at the viewer. Image generators often experience “pose bleeding” from reference uploads.

To force the model to break eye contact with the camera and lock onto the goblin, we must use aggressive directional keywords like “complete profile view”, explicitly mention his “turned head”, and move any mention of the viewer to a strict negative constraint

Here is a highly aggressive re-engineering of the Focused Archer prompt:

Aggressive Focused Archer Prompt

PURPOSE:
A narrative fantasy character painting in the traditional Dragonlance art style, depicting a halfling archer looking away from the camera to focus entirely on a target.

SUBJECTS AND ACTION:

  • Primary Subject: A 20-year-old male halfling archer. His exact facial features, hair type, and clothing style must be adapted from reference image @image1 . He stands 3’6″ tall, stout, with hobbit proportions, long black hair, a stubble moustache, a brown Sicilian hat, a green shirt, brown pants, and leather boots. He is chewing a strand of wheat.
  • Secondary Subject: A distant goblin standing nervously in the background, frozen in fear with a bright red apple balanced on its head.
  • Interacting Poses: The halfling is standing in a strict side-profile pose, completely turned away from the camera view. His head is turned sharply to the side, and his eyes are narrowed and locked entirely onto the distant apple. He is actively drawing his shortbow, pointing the arrow directly at the distant goblin. A mischievous, confident smirk remains on his face, but his gaze never leaves the target.

COMPOSITION:
A crisp side-profile or over-the-shoulder medium-wide shot. The camera is positioned behind or to the side of the halfling archer in the foreground, creating a straight visual line from his eye, down the length of the drawn arrow, directly to the nervous goblin in the background.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. Rich, earthy color palette. Bright, natural outdoor sunlight casting soft shadows.

CONSTRAINTS AND PRESERVATIONS:
Adapt the likeness from @image1 but heavily modify the pose. The character must NOT look at the camera. He must NOT face the viewer. Eye contact with the viewer is completely forbidden. His entire body, head, gaze, and bow arrow must be pointed and aimed directly at the goblin in the background. No modern digital art styles.

Why This Will Force the Look:

  • “Strict side-profile pose” / “completely turned away”: Repeating this across the Subject and Composition blocks forces the AI to draw the side of his face instead of copying the front-facing reference pose.
  • “Straight visual line”: Describing the camera tracking down his eyesight and arrow line forces the composition to connect the two characters rather than focusing on the viewer.
  • Forbidden Constraints: Moving “eye contact with the viewer is completely forbidden” into the constraints block tells the model’s negative prompt weights to actively suppress front-facing gazes.

Give this updated structure a spin. If it still clips his eyes back toward the front, we might need to describe his eye position mechanically (e.g., “only the profile of his left eye is visible as he looks down the arrow”). Let me know how the generator handles this version!

Notice here again how aggressively its pushing the DONT DO THIS parts. Its got the ” CONSTRAINTS” section filled with “must NOT” and “completely forbidden”. So the real question is… did all that work?

And the answer is… yes! So now lets *really* push it. Say I want to make it an action shot of him actually firing the arrow. “How would I turn this into a freeze-frame action shot of the moment the arrow hits the apple? Goblin looking up frightened. Halfling smirking in satisfaction. blurred motion around the arrow hitting the apple.”

This one took a few more iterations, with requests like “I want the same perspective.” and “I want him in the pose of having just release the arrow, and the arrow just making contact with the apple“. But I still ended up – with just 5 total renders from the start up top – with this prompt and result:

PURPOSE:
A high-tension narrative fantasy action painting in the traditional Dragonlance art style, capturing the precise split-second freeze-frame immediately after an arrow is released, viewed from behind the archer.

COMPOSITION AND PERSPECTIVE:
A deep-focus, wide environmental shot tracking down an outdoor shooting range. The camera is positioned behind the halfling archer’s right shoulder, creating a three-quarter rear view of the archer. The viewer looks over his shoulder, down the length of his extended left arm, straight downrange toward a goblin standing far in the distant background. Both the foreground archer’s back/profile and the distant target are in crisp, sharp focus simultaneously.

SUBJECTS AND ACTION:

Foreground Subject (Rear-View Follow-Through): A 20-year-old male halfling archer. He is facing completely away from the camera, looking downrange. His head is turned toward the target so only the profile of his cheek and jaw are visible from behind. His body and clothing style are strictly adapted from reference image @image1 , featuring his brown Sicilian hat, long black hair flowing down his neck, and leather armor. He is frozen in a post-release archery stance: his left arm is locked straight out toward the target holding his shortbow, his right hand is pulled back near his ear, and the empty bowstring is visibly vibrating.

Distant Background Subject: Positioned far down the range, a small but perfectly detailed goblin is looking upward in absolute fright and shock.

The Micro-Moment Impact: On top of the distant goblin’s head, a bright red apple is captured at the exact micro-moment of contact. The arrowhead has just slammed into and pierced the center of the apple, causing the fruit to fracture and split open.

Motion Blur: A linear streak of intense, localized motion blur bridges the entire distance, trailing directly behind the speeding arrow shaft from the foreground bow all the way to the distant target.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. High-speed freeze-frame clarity combined with painted motion streaks along the arrow’s trajectory. Bright, natural outdoor sunlight illuminates the field.

CONSTRAINTS AND PRESERVATIONS:
Adapt the character design from @image1 but heavily modify the orientation. The character must NOT face the camera. He must NOT look at the viewer. Eye contact with the viewer is completely forbidden. Only the back of his head, his shoulder, and his side-profile jawline are visible. The arrow must be hitting the distant apple. No modern digital art filters.

And that’s pretty awesome! The bowstring is on the wrong side of his lead arm, but we could probably fix that with some more iterations. Instead lets break it entirely – and ask it to switch him to a crossbow! “Switch him to firing a crossbow instead“

To switch from a regular bow to a crossbow, we need to completely overhaul the physical mechanics of his follow-through pose.

With a standard bow, the arms expand after release; with a crossbow, the weapon stays firmly braced against the shoulder, but the bowstring snaps forward violently against the front limbs. We must also swap out the “arrow” for a “bolt” or “quarrel”, as that is the correct terminology for crossbow ammunition.

Crossbow Release Impact Prompt

PURPOSE:
A high-tension narrative fantasy action painting in the traditional Dragonlance art style, capturing the precise split-second freeze-frame immediately after a crossbow is fired from behind the shoulder.

COMPOSITION AND PERSPECTIVE:
A deep-focus, wide environmental shot tracking down an outdoor shooting range. The camera is positioned behind the halfling archer’s right shoulder, creating a three-quarter rear view of the marksman. The viewer looks over his shoulder, directly down the wooden stock of the crossbow toward a goblin standing far in the distant background. Both the foreground crossbow follow-through and the distant target are in crisp, sharp focus simultaneously.

SUBJECTS AND ACTION:

  • Foreground Subject (Crossbow Follow-Through): A 20-year-old male halfling arbalest. He is facing completely away from the camera, looking downrange. His body and clothing style are strictly adapted from reference image @image1, featuring his brown Sicilian hat, long black hair flowing down his neck, and leather armor. He is frozen in a post-fire stance: his crossbow is held firmly braced against his shoulder, pointed straight downrange. The heavy steel bowstring of the crossbow has just snapped violently forward, depicted with vibrating, double-line visual resonance against the bow’s front limbs.
  • Distant Background Subject: Positioned far down the range, a small but perfectly detailed goblin is looking upward in absolute fright and shock.
  • The Micro-Moment Impact: On top of the distant goblin’s head, a bright red apple is captured at the exact micro-moment of contact. A short wooden crossbow bolt (quarrel) has just slammed into and pierced the center of the apple, causing the fruit to fracture and split open.
  • Motion Blur: A linear streak of intense, localized motion blur bridges the entire distance, trailing directly behind the speeding crossbow bolt from the foreground weapon all the way to the distant target.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. High-speed freeze-frame clarity combined with painted motion streaks along the bolt’s trajectory. Bright, natural outdoor sunlight illuminates the field.

CONSTRAINTS AND PRESERVATIONS:
Adapt the character design from @image1 but heavily modify the weapon and orientation. The character must NOT face the camera. He must NOT look at the viewer. The weapon must explicitly be a fantasy crossbow, not a shortbow. The projectile must be a short crossbow bolt hitting the distant apple. No modern digital art filters.

Key Crossbow Adjustments Made:

  • Braced Mechanics: Swapped out the old hand-at-jaw bow pose for a firm, over-the-shoulder braced stance, matching how a real crossbow is fired.
  • String Snap Reality: Clarified that the vibrating string is now resting against the front limbs of the weapon, which mathematically signals to the AI that the weapon was just triggered.
  • Ammunition Terminology: Changed “arrow” to “crossbow bolt (quarrel)” across all sections so the generator doesn’t accidentally draw a long, traditional arrow flying out of a mechanical weapon.

Not bad.. but the crossbow is missing some rather import pieces! So *this time* I pasted the image back into the chat and said “This is the result. Mostly good, but the crossbow is missing its bow.”

The image generation turned out beautifully—the perspective, the lighting, and the terrified goblin are absolutely spot-on.

The reason the crossbow is missing its horizontal bow arms (called the prod) is that you mentioned the bowstring snapping forward against the front limbs, but you also mentioned the shortbow from the previous prompt being completely swapped out. The AI got confused by the mechanical housing and left out the actual arched bow arms.

To fix this with an image editor, an inpainting brush tool, or by rerunning the prompt, we need to describe the weapon’s structural mechanics with explicit emphasis on the horizontal steel bow arms.

Here is the exact prompt adjustment to make sure the crossbow generates with a proper, complete bow assembly:

Fixed Crossbow Assembly Prompt

PURPOSE:
A high-tension narrative fantasy action painting in the traditional Dragonlance art style, capturing the precise split-second freeze-frame immediately after a heavy medieval crossbow is fired, viewed from behind the shoulder.

COMPOSITION AND PERSPECTIVE:
A deep-focus, wide environmental shot tracking down an outdoor shooting range. The camera sits behind the halfling’s right shoulder in a three-quarter rear view. The line of sight tracks straight down the wooden stock of the crossbow toward a terrified goblin far in the distant background. Both the foreground weapon and the distant target are in crisp, sharp focus simultaneously.

SUBJECTS AND ACTION:

  • Foreground Subject (Crossbow with Complete Bow Assembly): A 20-year-old male halfling arbalest facing completely away from the camera. His clothing and brown Sicilian hat are adapted from reference image @upload-1790603142405-2s3iw4x. He holds a heavy, complete fantasy crossbow braced firmly against his shoulder. Crucially, the crossbow features a prominent, wide horizontal steel bow (prod) mounted securely to the front of its wooden stock. The heavy steel bowstring has just snapped violently forward and is vibrating with double-line visual resonance, resting fully extended between the two tips of the horizontal bow arms.
  • Distant Background Subject: Positioned far down the range, a small but perfectly detailed goblin is looking upward in absolute fright and shock.
  • The Micro-Moment Impact: On top of the distant goblin’s head, a bright red apple is captured fracturing and splitting open as a short wooden crossbow bolt (quarrel) pierces right through the center.
  • Motion Blur: A linear streak of intense, localized motion blur trails directly behind the speeding crossbow bolt, bridging the space from the foreground bow tips all the way to the distant target.

STYLE AND LIGHTING:
Classic 1980s epic fantasy oil painting aesthetic matching the Dragonlance art style. High-speed freeze-frame clarity on the scene, with bright outdoor sunlight reflecting off the metallic surface of the crossbow’s steel bow arms.

CONSTRAINTS AND PRESERVATIONS:
Adapt the character design from @image1 but heavily modify the weapon orientation. The character must NOT face the camera. The weapon must be a complete mechanical crossbow, explicitly including its horizontal arched bow limbs extending out to the left and right sides. It must not look like a rifle or a stock without a bow. No modern digital art filters.

Key Crossbow Adjustments Made:

  • Braced Mechanics: Swapped out the old hand-at-jaw bow pose for a firm, over-the-shoulder braced stance, matching how a real crossbow is fired.
  • String Snap Reality: Clarified that the vibrating string is now resting against the front limbs of the weapon, which mathematically signals to the AI that the weapon was just triggered.
  • Ammunition Terminology: Changed “arrow” to “crossbow bolt (quarrel)” across all sections so the generator doesn’t accidentally draw a long, traditional arrow flying out of a mechanical weapon.

Now how’s *that* for an image that we once considered impossible!!


Leave a Reply