Hand an image editor a photo of an ordinary object, ask it for a face, and you get one of two outcomes: a drawn-on character that still reads as your original photo, or a repainted object that no longer matches the shot you started from. The difference sits almost entirely in how the instruction is worded. What follows is the prompt pattern for anthropomorphizing objects, the four things that pattern has to contain, and how to keep a whole set of objects in one style.
Anthropomorphism is a preservation problem
Adding eyes and limbs to an inanimate subject looks like a generation task and behaves like a preservation task. You want the model to hold the object's silhouette, texture, lighting and background exactly where they are, then place two or three new elements on top. A text-to-image model cannot do that. It starts from noise on every run and will hand you a different nut, a different table and a different key light along with the face you asked for.
Instruction-following image editors are built for this. They take an input image plus a natural-language edit, encode the image, and apply a localized change while leaving unaddressed regions close to untouched. Google's Nano Banana Pro is a common choice for this class of edit. ByteDance positioned Seedream 4.5, announced in December 2025, around the same two strengths: image editing and adherence to reference images. If you want open weights on your own hardware, FLUX Kontext from Black Forest Labs and Alibaba's Qwen-Image-Edit follow the same image-plus-instruction pattern.
The base prompt
This one comes from Kris (@icreatelife), who applied it to a photo of two nuts on a table:
Add illustrated googly eyes and stick hands on the chestnut and hazelnut in an exaggerated comedic pose
It is a good starting point because it is short and still fills every slot the model needs. Twenty words, four instructions, no wasted adjectives. Drop any one of the four and the output degrades in a predictable way.
The four slots
| Slot | What it does | Example wording |
|---|---|---|
| Rendering style | Tells the model the additions are drawn on, not sculpted from the material |
illustrated, flat vector, hand-drawn marker
|
| Features to add | Bounds the edit so nothing else changes | googly eyes and stick hands |
| Anchor | Names which object gets the treatment | on the chestnut and hazelnut |
| Pose or attitude | Gives the model a reason to place limbs somewhere specific | in an exaggerated comedic pose |
The slot people leave out most often is rendering style, and it is the one that does the most damage. With no style cue, the model tries to integrate the features into the object's own material: the eyes come back as glossy resin spheres embedded in the shell, correctly lit and deeply unsettling. The word illustrated is what keeps the additions reading as a layer on top of a photograph.
The anchor matters as soon as the frame contains more than one candidate object. Without it, an editor will often distribute features across everything in shot, or pick the largest object rather than the one you meant.
Variations
Same four slots, different content. This one adds an explicit preservation clause, which is worth including whenever the background is doing work:
Add flat cartoon eyes and thin noodle arms to the coffee mug, arms crossed, mildly annoyed expression, keep the original lighting and background unchanged
And one for a group, where the constraint is that the drawing style has to stay identical while the expressions differ:
Add hand-drawn marker eyes, eyebrows and stick legs to each of the three books on the shelf, a different expression on each, same drawing style across all three
Both run on any editor that accepts a reference image and an instruction. The second is a reasonable capability test for an editor you have not used before: it exercises localized editing, multi-object grounding, style consistency and preservation in a single pass.
Keeping a set consistent
One image is easy. A set of eight for a product page is where style drift shows up: the eyes get bigger, the line weight changes, one item comes back in a different illustration idiom entirely. A workflow that holds up:
- Produce the first image and treat it as the style reference for everything that follows, not just as output.
- On editors that accept multiple reference images, pass that first result alongside the new photo and ask for the same drawing style.
- Reuse the style wording verbatim across every prompt. Rewriting
illustrated googly eyesascartoon eyesmid-set is enough to shift the look. - Edit one object per pass. Batching four objects into one instruction reliably gives one good result and three approximations.
- When one item comes out wrong, re-edit that item rather than regenerating the set. Fix locally, keep the rest.
Failure modes
- The whole object gets repainted. Add an explicit preservation clause: keep everything except the added elements unchanged. Some editors respond better to naming what must stay than to naming what must change.
-
Features rendered as 3D objects. Push harder on the style slot:
2D,drawn on top,sticker,flat. Stack two of these if one is not enough. -
Features land on the wrong object. Anchor by material or position rather than by name, for example
the ceramic one on the left. - Additions too small to read. Give a relative size: eyes covering the top third of the object, arms as long as the object is wide.
- Expression ignored. Expressions are carried mostly by eyebrows and mouth shape. If you only asked for eyes, there is nothing for the model to act on. Name the parts that carry the emotion.
Where it earns its keep
The obvious use is social and product content: giving a physical product a face for a campaign without a photo shoot or an illustrator. It also works for explainer graphics, where turning an abstract component into a character makes a diagram easier to follow, and for mascot exploration, where you want twenty cheap directions before committing to one that gets drawn properly.
The technique is not limited to eyes and arms. Anything that reads as a drawn overlay behaves the same way: speech bubbles, motion lines, hand-drawn arrows, sketched annotations. The four slots are the same, only the feature list changes.
Takeaways
- Use an image editor, not a text-to-image model. This is a localized edit on an existing photo.
- Every prompt needs rendering style, features, anchor and pose. The style cue is the one that prevents uncanny results.
- Add an explicit preservation clause when the background matters.
- For sets, lock the style wording, edit one object at a time, and feed the first accepted result back as a reference.
Top comments (0)