Your image technically has everything you asked for — the cat, the coffee cup, the book, the laptop, the window, the houseplant, the morning light — but it reads as visual chaos. The eye doesn’t know where to land. Every object is rendered at similar size, sharpness, and prominence, so the brain reads “noise” instead of “scene.”
Cluttered composition is rarely a “the model can’t compose” problem. It’s almost always a prompt problem: you listed seven things and gave the model no priority signal.
TL;DR — fastest fix
Rewrite the prompt around one hero subject plus at most two secondary objects, then add a single depth-of-field line so everything else falls out of focus:
a ginger cat sitting on a desk, sharp focus,
soft morning window light, one out-of-focus coffee cup beside it,
shallow depth of field, f/1.4, creamy bokeh, minimalist compositionThat one rewrite clears the large majority of clutter. If you already have a finished image you like and just want to remove one or two intruding objects, skip the re-roll and use the region/inpaint editor (Step 7 below) — Midjourney’s Vary (Region), ChatGPT’s selection edit, or Gemini’s region edit. Both routes are detailed below.
Which bucket are you in?
| Symptom in the output | Most likely cause | Go to |
|---|---|---|
| 5+ objects all at similar size, all sharp | Too many equal-weight nouns | Cause 1, Steps 1-2 |
| Background props as crisp as the subject | No depth-of-field cue | Cause 2, Step 3 |
| Several things “could” be the subject | No explicit hero | Cause 3, Step 2 |
| Frame packed wall-to-wall with stuff | Wide framing + scene words | Cause 4, Steps 1, 4 |
| Style itself looks busy (flat lay, still life) | Style anchor implies clutter | Cause 5 |
| One finished image, just 1-2 intruders to remove | Don’t re-roll | Step 7 |
Common causes
Ordered by hit rate, highest first.
1. Too many objects with equal weight in the prompt
cat, coffee, book, laptop, plant, window light, cozy morning — seven nouns, no hierarchy. The model treats them as equally important and tries to render all of them at central prominence.
How to spot it: count the concrete nouns in your prompt. More than 3 without weighting and cluttered output is likely.
2. No depth-of-field cue
Without DOF instructions, the model defaults to a deep, everything-in-focus look. Peripheral elements then compete with the subject for attention because nothing visually recedes.
How to spot it: your prompt has no shallow depth of field, bokeh, f/1.4, out of focus, or blurred background. Add one.
3. No explicit hero subject
You said the cat is in the scene, but you didn’t say the cat is the subject. Models need that hierarchy hint, especially when multiple nouns are listed.
How to spot it: your prompt doesn’t have hero subject, main subject, centered, dominant, or a sized modifier like large cat, tiny coffee cup in the background.
4. Wide framing with detailed scene words
Wide shot plus words like cozy, interior, room, still life, lifestyle scene invite the model to fill the frame with stuff. Tighter framing or single-noun composition prevents it.
How to spot it: prompt is wide and uses scene / lifestyle / interior-style words.
5. Style anchor implies clutter
Specific styles bake in clutter:
still life painting— multiple objects on a tablecozy aesthetic— many props, soft layered detailflat lay photography— busy by definitionwes anderson— symmetrical maximalismstudio ghibli interior— busy lived-in spaces
How to spot it: your style anchor evokes a busy scene on its own.
Shortest path to fix
Step 1: Cut to one hero plus at most 2 secondary objects
Before:
a cat, a coffee cup, a book, a laptop, a houseplant, a window with morning light, a cozy desk sceneAfter:
a ginger cat sitting on a desk, soft morning window light in the background,
one out-of-focus coffee cup beside the catOne hero (cat), one secondary object (coffee cup, explicitly out-of-focus), and atmosphere (window light) instead of yet another object.
Step 2: Add an explicit hero subject plus size modifiers
Prompt patterns that work:
"[hero] is the main subject, centered, large in frame"
"close-up of [hero], everything else small and out of focus"
"[hero] in sharp focus, [other objects] blurred in the background"Step 3: Add depth of field
This single line turns most “everything is sharp” cluttered images into “subject pops”:
"shallow depth of field, f/1.4, creamy bokeh, only [hero] in focus"Step 4: Add negative-space wording
Words to add (pick 1-2):
minimalist compositionlarge negative spacebreathing room around the subjectclean composition with simple backgroundJapanese minimalist aesthetic(if it fits your style)
Step 5: Tool-specific clutter controls
Midjourney (V8.1 default as of June 10, 2026; V7 still selectable with --v 7). Append:
... --style raw --ar 4:5 --s 100--style raw strips Midjourney’s automatic “beautify” pass, which is a major source of unrequested extra detail. --ar 4:5 (a tall ratio added in the V8 range, which supports anything from 1:2 to 2:1) shrinks the background area the model can fill. --s (stylize) defaults to about 100; lowering it toward --s 50 or --s 0 keeps the model closer to your literal prompt and adds less flourish. Confirm the current parameter behavior in the Midjourney Parameter List.
ChatGPT (Images 2.0 / gpt-image-2, released April 21, 2026). There are no -- flags. Put the hierarchy in plain language and follow up conversationally: “Make the cat the clear hero, blur and shrink everything else, lots of empty space around it.” Supported aspect ratios run 3:1 to 1:3, so ask for a portrait frame to cut background room.
Gemini (Nano Banana 2, launched February 2026). Same plain-language approach; it follows instruction-style edits well, so a turn like “remove the laptop and the plant, keep only the cat and the cup, blur the background” usually lands in one pass.
Step 6: Negative-prompt the clutter (SD / SDXL family only)
In Stable Diffusion, SDXL, ComfyUI, or Forge, add to the negative prompt:
cluttered, busy composition, many objects, crowded scene,
multiple subjects, ornate, baroque, maximalist, busy background,
overlapping objectsNegative prompts only exist in the Stable Diffusion family. Midjourney’s equivalent is --no (for example --no laptop, plant); ChatGPT and Gemini have no negative-prompt field, so phrase exclusions in plain language instead.
Step 7: Remove intruders post-hoc instead of re-rolling
If a single generation is otherwise great and only one or two objects ruin it, edit in place rather than regenerating the whole image:
- Midjourney Editor (web) — open the image, choose Vary (Region), draw a box or lasso over the unwanted object, and run it. Midjourney recommends selecting roughly 20-50% of the image so it has enough context. To delete rather than replace an object, type
empty(or a plain background description) as the region prompt. See the Midjourney Editor docs. - ChatGPT Images 2.0 — use the selection tool to highlight the object, then say “remove this and fill with the background.” It keeps everything else untouched.
- Gemini Nano Banana 2 — describe the edit directly: “remove the book on the left, blend the desk surface behind it.”
Step 8: Sketch the composition first, then lock it (ControlNet)
For art-directed work, do a rough sketch (paper, iPad, or ControlNet’s Scribble) of the composition you want, then feed it as a ControlNet input. The model fills in detail but cannot deviate from your layout.
# ComfyUI / Forge ControlNet (SDXL)
- Load ControlNet Scribble or Canny (the SDXL ControlNet Union model covers both)
- Provide your composition sketch as the control image
- Control weight (strength): start at 0.6-0.7
(lower = more model creativity, higher = stricter adherence)
- If stacking two ControlNets, keep each at 0.5-0.7 so neither dominatesHow to confirm it’s fixed
You’ve actually solved it, not just shuffled the clutter, when:
- You can name the hero subject in one second when you glance at the thumbnail.
- The background is visibly softer or simpler than the subject (real depth separation, not just darker).
- There are no fewer than two areas of clear negative space the eye can rest in.
- Removing any one remaining object would not change “what the picture is about.”
If the thumbnail (not the full-size image) still reads as busy, you haven’t fixed the composition — the eye sorts a scene at thumbnail scale first.
Prevention
- Decide the hero subject BEFORE writing the prompt; write it first in the sentence.
- Default to 3 nouns max per image; if you need more, make a series, not one image.
- Set a rule: every prompt with 3+ nouns must include a depth-of-field or focus modifier.
- For series work, keep a reusable “minimalist composition, shallow depth of field” snippet at the end of every prompt.
FAQ
Why does the model keep adding objects I never asked for?
Two reasons. First, busy style anchors (cozy, flat lay, still life, lived-in interior) carry implied props, so the model adds them to satisfy the style. Second, on Midjourney the default stylize pass invents extra detail; --style raw and a lower --s value rein that in. Strip the style word or switch to raw mode and the phantom objects usually disappear.
Does a negative prompt work in Midjourney, ChatGPT, or Gemini?
No. A true negative-prompt field exists only in the Stable Diffusion / SDXL family. Midjourney uses --no object for exclusions; ChatGPT Images 2.0 and Gemini have no exclusion field at all, so you state what to leave out in plain language (“no laptop, no plant in frame”).
My image is otherwise perfect but has one extra object. Do I have to regenerate?
No, and you shouldn’t — a re-roll changes everything. Use region editing instead: Midjourney’s Vary (Region) with empty as the prompt, ChatGPT’s selection edit, or a Gemini “remove this object” turn. All three keep the rest of the image intact.
I added shallow depth of field but everything is still sharp. What now?
Some models weakly honor abstract DOF terms. Make it concrete and stack cues: name an aperture (f/1.4), add creamy bokeh, and explicitly say which element stays sharp (only the cat in focus, background blurred). On Midjourney, a longer-lens cue like 85mm portrait lens also pushes the background out of focus.
How many objects is too many? As a working rule, 3 concrete nouns is the ceiling for a clean single image, and that already assumes you weight them (one hero, the rest secondary and out of focus). Beyond 3, plan a series or a collage rather than forcing one frame to carry all of them.