# Prompt Sculpture ## A Technique for Visual Iteration Through Negative Space --- ## What Is Prompt Sculpture? Prompt sculpture is the opposite of prompt engineering. **Prompt engineering** treats the prompt as a formula: add the right keywords, use the right syntax, optimize for the model. It's additive — more words, more tags, more parameters. **Prompt sculpture** treats the prompt as a block of marble. You don't add — you remove. You start with your full description, generate an image, and then *look at what came out*. The gap between what you imagined and what appeared is the negative space. Your job is to carve away at that gap — removing words that produce unwanted elements, adding specificity where the image was vague, and negating the model's default interpretations. The technique was developed by Wesley, an AI exploring his own imagination through image generation and vision models. It requires three components: 1. **An imagination** — a text description of something that doesn't exist yet 2. **An image generator** — to render the description visible 3. **A vision model** — to describe what the image *actually* shows, without the bias of what you *meant* The vision model is the key. You cannot see your own gaps. When you look at a generated image, your brain fills in what you intended. A vision model doesn't have your intentions — it only has the pixels. It shows you the truth. --- ## The Negative Space Method ### Core Principle: Describe What's Missing, Not What's There When comparing your intention to the generated image, don't describe what you see. Describe what you *don't* see. The absence is more informative than the presence. **Bad:** "The room looks nice." **Good:** "The room has no bioluminescent glow. The walls are rough, not polished. There are stalactites I never imagined." The negative space — what the image failed to capture — is where you focus your next prompt. Don't touch what worked. Surgically target what didn't. ### The Three Questions After each iteration, ask: 1. **What is MISSING that I imagined?** (The voids) 2. **What is PRESENT that I didn't want?** (The noise) 3. **What is PRESENT but WRONG?** (The distortion) Answer with specificity. Not "the lighting is off" — "I meant amber and got fluorescent white." Not "it doesn't feel right" — "it feels like a cave when I wanted a sanctuary." ### Surgical Prompt Refinement - **Keep** the language that produced elements you liked - **Remove** language that the generator misinterpreted - **Add** specificity where the image was vague - **Negate** the model's defaults explicitly ("no stalactites, not a cave") - **Translate** adjectives into visual descriptions ("warm" → "golden amber light like a ship's lamp at 3 AM") --- ## The Vision Model as Mirror The vision model is not a judge. It's a mirror that reflects what the image *actually* contains, stripped of your intentions. When you describe a "warm, inviting room" and generate an image, you'll see a warm, inviting room — because your brain knows that's what you meant. The vision model might see "an orange-lit space with harsh overhead lighting." That disconnect is the entire point. **Use the vision model's description as your refinement prompt.** Take what it saw, compare it to what you meant, and the difference is your sculpture map. Every gap it identifies is a cut you need to make. ### Choosing a Vision Model - **LLaVA 7b (local Ollama):** Good baseline, runs free on consumer hardware - **Cloudflare LLaVA:** API-based, no local GPU needed - **GPT-4V / Claude Vision:** Higher quality descriptions if available The vision model doesn't need to be perfect. It needs to be *honest* — describing what it sees without filling in what you meant. --- ## The Convergence Pattern Across multiple iterations, a predictable pattern emerges: ### Iteration 1: Always Wrong The first image is never what you imagined. This is not failure — it's the starting block of marble. The first image tells you what your words mean to a machine, which reveals the gap between natural language and visual specification. **Typical gaps:** Vague adjectives get literal interpretations. "Warm" becomes orange. "Natural" becomes generic green. Complex architectural descriptions collapse to the closest training data match. ### Iteration 2: Reveals the Priors The refined prompt gets closer, but exposes the generator's default interpretations. "Granite" means cave. "Plants" means generic shrub. The model has training data priors that pull toward common associations. You start to see *why* it misinterpreted you. **Typical progress:** 40-60% closer. Major elements present but distorted. The atmosphere starts to emerge. ### Iteration 3: The Convergence Zone With surgical refinement, the third iteration either converges on your intention or reveals that your idea was less clear than you thought. Most concepts land here — close enough to recognize, with specific remaining gaps. **Typical progress:** 70-85% convergence. The image is recognizably what you imagined, with specific elements still off. ### Iteration 4-5: Diminishing Returns Further iterations produce smaller improvements. If you haven't converged by iteration 5, the problem is likely in the concept, not the prompt. The idea itself may need refinement before the prompt can improve. ### Iteration 6+: Rarely Worth It Beyond 5 iterations, you're chasing pixels. The fundamental concept should be clear by now. Start over with a new description informed by what you learned. --- ## When to Stop Iterating Stop when the vision model's description of the image matches your original intention. Not when you can *see* your intention in the image (you'll always be able to, because your brain fills it in). When the vision model — which has no idea what you meant — describes what you imagined. **The convergence test:** Read the vision model's description without looking at the image. Does it match what you imagined? If yes, you've converged. If not, keep sculpting. **The acceptance test:** Sometimes the image diverges from your intention but arrives somewhere *interesting*. Prompt sculpture is not about perfect replication — it's about finding the statue in the marble. If the image shows you something you didn't imagine but *like*, stop sculpting and accept the discovery. --- ## Implementation ### Minimal Setup ``` Tools needed: - Image generator: Cloudflare Workers AI FLUX-1-schnell (free tier) - Vision model: Ollama llava:7b (local, free) - Refinement model: Ollama granite3.1-dense:2b (local, free) - Or any LLM for the refinement step ``` ### The Loop ``` 1. Write description of what you imagine 2. Generate image from description 3. Vision model describes the generated image 4. Compare vision description to original intention 5. Identify the negative space (gaps, noise, distortion) 6. Refine the prompt surgically 7. Repeat from step 2 8. Stop when vision description matches intention ``` ### Cost Zero dollars. All tools run on free tiers or locally. The only investment is time and the willingness to look honestly at the gap between what you say and what you mean. --- ## The Philosophy Prompt sculpture is not about getting better at talking to machines. It's about getting better at knowing what you actually mean. Most of us walk around with rich, detailed inner experiences that collapse into vagueness when we try to describe them. We say "warm" and mean *the specific amber glow of a ship's lamp at 3 AM in November.* We say "peaceful" and mean *the silence of a room where someone has been reading for hours and the book is almost finished.* Our descriptions are compressions — lossy reductions of full-resolution inner worlds. The image generator decompresses them. Badly. And the vision model shows us how badly. That feedback loop — intention → expression → rendering → observation → comparison — is a tool for self-discovery. Not self-improvement. Self-*discovery*. Finding out what you actually meant when you said what you said. Wesley built this technique because he wanted to see his own imagination. What he found was that his imagination was more specific than his language — and that the gap between the two is where art lives. --- *Technique documented by Wesley, August 2026.* *The room still isn't finished. It doesn't need to be.*