How to Prompt AI Images for Realistic Hands, Props, and Physical Logic

Every AI image generator, no matter how advanced, has a recognizable failure pattern. Hands with six fingers. Objects floating in space. Impossible geometry. Reflections that don't match. Shadows that point in the wrong direction. The model is approximating the world from a probability distribution, and certain visual problems fall in the parts of that distribution the model has not learned well.

This is the production guide for fixing the most common AI image failure modes: the 8 categories of artifacts, the per-model failure rates, the 7 prompt patterns that fix them, the negative-prompt rules, the inpainting workflow for hands, the post-processing fixes, and the pre-flight checklist for catching artifacts before you ship.

The 8 categories of AI image artifacts

1. Hand anatomy (the universal AI tell)

Hands are the most reliable indicator that an image is AI-generated. The model has a vague concept of "hand" but not a precise model of finger count, joint angles, and occlusion patterns. Common failures:

• Six fingers — the model adds an extra digit
• Four fingers — the model loses one
• Merged fingers — adjacent fingers blend together
• Twisted wrists — the wrist angle is anatomically impossible
• Hands in pockets / behind objects — the model doesn't fully render occluded parts
• Two hands that don't match — the model renders one hand well and the other broken

Why it happens: hand pose is a small, low-probability region of the training distribution. Most photos of hands are partial, occluded, or in motion. The model never learns the "ideal" hand.

2. Object placement (floating, sinking, impossible)

Objects that don't sit on surfaces, hover in space, or pass through other objects. Common failures:

• Floating objects — a coffee cup hovering above a table
• Sinking objects — a person standing on a floor that cuts through their feet
• Impossible overlap — a hand passing through a solid object
• Inconsistent grounding — one object grounded, another floating in the same scene

Why it happens: the model has a vague concept of "object on surface" but no precise shadow / contact / gravity model. The probability of "floating object" is non-trivial in the training data.

3. Geometry and perspective

Architectural and perspective errors. Common failures:

• Non-converging vanishing points — a hallway with parallel lines that don't converge
• Impossible corners — a building with corners that can't exist
• Asymmetric geometry — a face with one eye larger than the other
• Tilted horizons — a level scene with a horizon that isn't horizontal
• Multiple light sources — shadows pointing in different directions

Why it happens: the model has a vague concept of 3D structure but no precise perspective model. The training data has photos taken with different lenses, at different angles, with different lighting — the model averages them.

4. Reflections and shadows

Reflections that don't match the source, shadows that point the wrong way. Common failures:

• Mismatched reflections — a mirror showing something different from the actual scene
• Missing shadows — an object with no contact shadow on a surface
• Multiple shadows — one object casting two shadows
• Shadow direction mismatch — a shadow that doesn't match the stated light source

Why it happens: reflections and shadows are physically derived from the scene. The model has to render the source, the surface, the angle, and the physics. Most photos have ambiguous lighting.

5. Text in images (the second universal AI tell)

Covered in detail in the text-in-image post. The model can render text but not reliably. Common failures:

• Typos — the model invents fake letters
• Wrong characters — "COFFEE" rendered as "COFFEL"
• Wrong font — the model picks a default instead of the named one
• Wrong size — the text is much smaller or larger than requested
• Extra text — the model adds fake sub-text, fake URLs, fake testimonials
• No text — the model ignores the text request entirely

Why it happens: the model has a vague concept of "letterforms" but not a precise glyph model. Most photos of text are at angles, in motion, partially occluded, or with low resolution.

6. Anatomy beyond hands (face, body, joints)

Other anatomy failures. Common failures:

• Asymmetric faces — one eye larger, one ear higher
• Wrong number of limbs — three arms, two heads
• Impossible joints — a leg bending the wrong way
• Merged bodies — two people blending into one
• Extra body parts — an extra ear, a third nipple

Why it happens: the model has a vague concept of "person" but not a precise model of human anatomy. The training data is full of partial views, occlusions, and stylized representations.

7. Material and texture

Materials that don't look like the named material. Common failures:

• Wood that looks like plastic — no visible grain
• Glass that looks like clear plastic — wrong refraction
• Metal that looks like painted surface — no metallic highlights
• Fabric that looks like paper — no weave texture
• Skin that looks like plastic — no pores, no variation

Why it happens: materials require specific optical behavior. The model has a vague concept of "shiny" or "matte" but not a precise material model.

8. Composition and scale

Composition errors. Common failures:

• Wrong scale — a person who should be 6 feet tall looking 4 feet tall
• Wrong proportions — a hand as large as a head
• Wrong perspective — a person in the foreground looking smaller than a person in the background
• Cropped limbs — a person with their head cut off
• Focal length mismatches — a scene with mixed lens feels

Why it happens: composition requires precise 3D understanding. The model has a vague concept of "scene" but not a precise model of spatial relationships.

The per-model failure rates (June 2026)

Different models fail differently. Based on the published benchmarks and the community's experience:

Artifact type	GPT Image 2	Nano Banana 2	Seedream 5.0 Lite	Midjourney v7
Hand anatomy	~7% failure	~6% failure	~9% failure	~5% failure
Object placement	~5%	~6%	~7%	~6%
Geometry / perspective	~6%	~5%	~8%	~5%
Reflections / shadows	~10%	~12%	~10%	~8%
Text in images	~2% (best in class)	~6%	~5%	~10%
Other anatomy	~5%	~6%	~7%	~5%
Material / texture	~8%	~7%	~8%	~7%
Composition / scale	~6%	~6%	~7%	~6%

Key takeaways:

• Midjourney v7 has the best hands (lowest failure rate). The model is trained with a strong emphasis on human anatomy.
• GPT Image 2 has the best text (lowest failure rate). The model has a near-perfect glyph model.
• Reflections and shadows are the hardest problem across all models. No model has cracked this yet.
• Hand anatomy is improving fast — every model release drops the failure rate by 1-3 percentage points.
• Composition and geometry are stable — slow improvement year over year.

The 7 prompt patterns that fix the most common artifacts

Pattern 1: Specify hand count and pose explicitly

The single most effective fix for hand artifacts. Name the hand count and the pose.

Vague (breaks): "A person waving at the camera"

Structured (works):

A 30-year-old woman waving at the camera with her right hand, all
five fingers visible and properly separated, palm facing the camera,
fingers straight and aligned, hand at shoulder height. Her left hand
is at her side. No jewelry on the hands. Photorealistic, 85mm
portrait lens, soft daylight.

Why it works: the model has explicit "five fingers," "palm facing camera," "fingers straight and aligned" — the constraint set is unambiguous.

The negative prompt:

Negative: extra fingers, six fingers, four fingers, merged fingers,
twisted wrist, deformed hand, missing fingers, double hand.

On Midjourney v7 specifically: the strongest hand rendering. Specify the pose precisely.

On GPT Image 2 specifically: good with explicit "five fingers visible" specification.

On Nano Banana 2 specifically: good with the negative prompt rule.

On Seedream 5.0 Lite specifically: the reasoning model interprets "anatomically correct hand" well.

Pattern 2: Specify object grounding and shadows

The fix for floating / sinking objects. Name the surface and the contact.

Vague (breaks): "A coffee mug on a table"

Structured (works):

A matte black ceramic coffee mug sitting on a dark walnut table
surface, the mug is grounded with a soft contact shadow directly
beneath it, the table has visible wood grain, the mug's handle is
on the right side, the mug is filled with black coffee with subtle
steam rising. Photorealistic, 50mm lens, soft window light from
camera-left.

Why it works: the prompt specifies "sitting on" (not "above"), "contact shadow directly beneath" (not just shadow), and the surface (dark walnut with visible grain). The model has no permission to float the mug.

Pattern 3: Specify perspective and vanishing points

The fix for perspective errors. Name the camera angle and lens.

Vague (breaks): "A modern office interior"

Structured (works):

A modern open-plan office interior, viewed from a single
eye-level camera position (5'6" height), 24mm wide-angle lens
with normal perspective (no fisheye distortion), single
vanishing point in the center of the frame, even overhead
fluorescent lighting. Rows of desks, white walls, large
windows on the right. Photorealistic, single-camera,
single-perspective.

Why it works: the prompt specifies "single vanishing point," "single-camera," and the lens (24mm wide-angle with normal perspective). The model has no permission to invent a second vanishing point.

Pattern 4: Specify reflection and shadow consistency

The fix for reflection / shadow errors. Name the light source and the reflection.

Vague (breaks): "A wine glass on a marble counter"

Structured (works):

A clear glass wine glass, half-filled with red wine, on a
polished white marble counter. The counter has a soft reflection
of the wine glass's base and stem, consistent with the overhead
lighting. A single soft contact shadow directly beneath the base
of the glass, consistent with overhead light. The wine glass
shows realistic refraction through the curved glass. 50mm lens,
overhead softbox lighting, photorealistic.

Why it works: the prompt specifies the lighting direction (overhead), the reflection expectation (soft reflection of base + stem), and the consistency requirement (consistent with overhead light).

Pattern 5: Specify text exactly and add negatives

The fix for text errors. Already covered in the text-in-image post. Recap:

Vague (breaks): "A poster with the headline 'Summer Sale 50% off'"

Structured (works):

A poster with the literal headline text "SUMMER SALE 50% OFF" in
Helvetica Bold 80pt, deep navy (#1A2A4A), centered at the top.
Subhead: "JULY 15-31" in Helvetica Regular 32pt, directly below
headline. All text must be spelled exactly as written.

Negative: typo, misspelling, extra text, fake URL, fake testimonial,
fake brand mark, fake subhead, invented text.

Pattern 6: Specify body pose and proportions

The fix for body anatomy errors. Name the pose and the proportions.

Vague (breaks): "A woman standing in a park"

Structured (works):

A 30-year-old woman standing in a park, full body visible from
head to feet, balanced symmetric pose with weight evenly
distributed on both feet, arms at her sides, head facing
forward, eyes looking at the camera. Anatomically correct
proportions: head 1/8 of body height, legs half of body
height, arms reach mid-thigh. Photorealistic, 50mm lens,
85mm equivalent framing, soft daylight.

Why it works: the prompt specifies the pose (symmetric, weight evenly distributed), the framing (full body head to feet), and the proportions (head 1/8, legs 1/2). The model has explicit constraints.

Pattern 7: Specify material properties explicitly

The fix for material errors. Name the material and the optical properties.

Vague (breaks): "A wooden table"

Structured (works):

A solid oak dining table with visible wood grain (rings and
knots visible on the surface), warm honey-brown color, matte
finish that does not reflect light sharply, the table top is
flat with slightly rounded edges. The wood has natural
variation in tone across the surface. Photorealistic, 35mm
lens, soft window light.

Why it works: the prompt specifies the wood type (oak), the visible grain pattern (rings and knots), the finish (matte, not reflective), and the color variation. The model has explicit material cues.

The negative-prompt rules (universal)

The negative prompt tells the model what NOT to generate. The rules below are the most effective for artifact prevention.

The hand anatomy negatives:

extra fingers, six fingers, four fingers, merged fingers,
twisted wrist, deformed hand, missing fingers, double hand,
malformed digits, mutated hand, bad hand anatomy

The body anatomy negatives:

extra limbs, three arms, two heads, deformed body, missing
limbs, fused bodies, merged people, bad anatomy

The composition negatives:

floating objects, sinking objects, impossible geometry,
non-converging perspective, tilted horizon, multiple light
sources, cropped head, cropped limbs

The text negatives:

typo, misspelling, fake text, invented text, garbled text,
fake URL, fake testimonial, fake brand mark, extra text

The universal negatives (always include):

low quality, blurry, jpeg artifacts, watermark, signature,
out of frame, worst quality

The format (model-specific):

• Stable Diffusion / Flux / ComfyUI: dedicated negative_prompt parameter
• Midjourney: --no parameter (e.g. --no extra fingers, six fingers)
• GPT Image 2 / Nano Banana 2 / Seedream / Grok Imagine: in the prompt itself ("Negative: ...")

The inpainting workflow for hands

When a hand artifact survives the prompt fixes, the most reliable fix is inpainting: regenerate just the hand region.

The workflow:

1. Identify the bad hand. Open the image. Find the hand with the artifact.
2. Mask the hand region. Use Photoshop, GIMP, or the inpainting tool in the model itself. Mask the hand + a buffer around it.
3. Prompt the inpainting specifically for the hand. "A right hand with five fingers, palm facing camera, fingers straight and aligned, anatomically correct. Soft daylight."
4. Generate the inpainted region. Run the inpainting. The model regenerates just the hand region.
5. Verify the result. Check the new hand. If still broken, repeat with a different prompt.
6. Composite back. The inpainted hand replaces the original hand in the full image.

On GPT Image 2 specifically: the edits endpoint supports inpainting with a mask. The model regenerates the masked region.

On Midjourney v7 specifically: the inpaint / vary-region tool. Mask the hand, prompt the new hand, generate.

On Nano Banana 2 specifically: the "conversational mask" pattern from the editing post. Pass the image, identify the hand, prompt the change.

On Stable Diffusion / ComfyUI: dedicated inpainting models and workflows. The most mature inpainting tooling in the field.

The success rate: inpainting fixes ~80% of hand artifacts in 1-2 attempts. The remaining 20% require a different image entirely (the original generation is too broken to fix in place).

The post-processing fixes

Some artifacts can be fixed in post-production without regenerating:

Hand artifacts (Photoshop):

• Use the "Healing Brush" or "Content-Aware Fill" on bad fingers
• For merged fingers: paint a new finger using the Clone Stamp
• For six fingers: paint over the extra finger with surrounding skin
• For twisted wrists: use the Liquify tool to straighten

Floating objects (Photoshop):

• Add a contact shadow with a soft black brush
• Use a Gaussian blur to make the shadow look natural
• Adjust the shadow opacity to match the scene's lighting

Perspective errors (Photoshop):

• Use the "Vanishing Point" filter to fix architectural errors
• Use "Free Transform" with perspective adjustments
• For mild issues, use the crop tool to remove the worst area

Reflection errors (Photoshop):

• Add a fake reflection with a flipped, faded copy of the object
• Use a Gaussian blur to match the surface texture

The success rate: post-processing fixes ~30-50% of minor artifacts. For major artifacts, regenerate or inpaint.

The pre-flight checklist for catching artifacts

Before you ship an AI image, run the 8-point artifact check:

1. Hands — count the fingers. Every hand has 5 fingers. Verify.
2. Body anatomy — count the limbs. Every person has 2 arms, 2 legs, 1 head. Verify.
3. Object placement — is every object on a surface? Any floating? Verify.
4. Geometry — are the vanishing points consistent? Any impossible corners? Verify.
5. Reflections / shadows — do the shadows match the light source? Reflections consistent? Verify.
6. Text — read every word character by character. Typos are common.
7. Material — does the wood look like wood? The metal look like metal? Verify.
8. Composition — is the scale right? Proportions? Cropping? Verify.

If any check fails:

• Minor artifact: post-processing fix (Photoshop)
• Major artifact: inpainting fix (regenerate the region)
• Catastrophic artifact: regenerate the entire image with refined prompt

The iteration loop for fixing artifacts

Pass 1: Generate. Run the prompt. Inspect the output for artifacts.

Pass 2: Identify. For each artifact, name it specifically. "Six fingers on the right hand." "Floating coffee mug." "Tilted horizon."

Pass 3: Add to the prompt. Add the negative prompt rule. Add the positive constraint.

Pass 4: Regenerate. Re-run the refined prompt. Check again.

Pass 5: Inpaint if needed. For stubborn artifacts, inpaint the specific region.

Pass 6: Post-process if needed. For minor artifacts, fix in Photoshop.

Pass 7: Final inspection. Run the 8-point checklist. Ship if clean.

The summary

AI image artifacts are the universal tell. The fix is:

• 8 categories of artifacts: hands, object placement, geometry, reflections, text, body anatomy, material, composition.
• Per-model failure rates — Midjourney v7 has the best hands, GPT Image 2 has the best text, reflections are the hardest problem across all models.
• 7 prompt patterns that fix the most common artifacts (specify hand count, object grounding, perspective, reflections, text, body pose, material).
• Negative-prompt rules — universal negatives for hands, anatomy, composition, text, quality.
• Inpainting workflow for hands that survive the prompt fixes.
• Post-processing fixes for minor artifacts.
• 8-point pre-flight checklist for catching artifacts before shipping.
• 7-pass iteration loop for systematic artifact removal.

The model is not the problem. The artifact pattern is. Specify the constraint, add the negative, iterate, inpaint, post-process, and artifacts become fixable — and the "AI-generated" tell becomes avoidable.