The Concept: Why AI Remodeling?
When we started developing AI Remodeling TV, the brief was simple: show the AI, don’t hide it. Most content that uses AI image generation tries to pass the output off as real, or buries the process. We wanted the opposite — the transformation itself is the content.
The question was what to transform. We landed on real places and things that people already have strong visual associations with. Walmart. Lincoln Memorial. A DC corner store. The stronger the source recognition, the more powerful the AI output. You need to know what it was to feel what it became.
🔴 The Rule We Built Around
The subject must remain recognizable after the transform. If you can’t tell it was a Walmart, the transform failed — regardless of how good the output looks. Familiarity is the whole point.
The Tools: What We Actually Used
We tested six AI image generation tools before locking in our production stack. Most of the discourse online is about which model is “best” — that’s the wrong question. Different models are better at different things, and our process runs them in parallel.
- Midjourney v6.1 — Best for photorealistic architectural output. The level of structural detail and lighting control is unmatched. We used this for every Known Places episode.
- DALL·E 3 — Best for color drama and abstract interpretations. Where Midjourney gives you realism, DALL·E gives you expressionism. Stronger for Viral Clips transforms.
- Adobe Firefly — Best for clean geometric output. Signage, type, logos, structured elements render better here than anywhere else. Used for any transform where branded elements need to stay legible.
- Stable Diffusion XL — We tested it but didn’t use it in Season 1. Control is high but output quality ceiling was below the others at our use case scale.
“Don’t ask which model is best. Ask what you’re trying to do, then pick the model that does that specific thing best.”
— Internal production note, Pressure Media
The Prompt System
Our prompts follow a rigid structure. Every ARTV prompt has six components in sequence:
- Subject anchor — Describe the original subject in plain language. Be specific. “Walmart Supercenter exterior” not “big store.”
- Transform style — The aesthetic direction. One clear style. Not “cyberpunk and brutalist and organic.” Pick one.
- Preservation rules — What must stay recognizable. Scale, proportion, specific elements. These are constraints, not suggestions.
- Atmosphere tags — Lighting, mood, weather, time of day. These carry more weight than most people realize.
- Quality specifiers — Cinematic, photorealistic, 8k, etc. These are last for a reason: they should never override the conceptual direction.
- Model parameters — –ar, –v, –style, –q, –s. Model-specific controls that tune output toward your intent.
⚠️What We Got Wrong Early
We over-specified transforms in early prompts. Five adjectives for the aesthetic direction produces muddy output. One strong direction produces a clear result. Specificity is in the preservation rules and atmosphere, not the style itself.
The Failures (Documented)
We ran approximately 340 prompts across Season 1 to produce 7 episodes. Most didn’t make the cut. The most common failure modes:
- Recognition collapse. The transform was too complete. Beautiful output, but you couldn’t tell it was the original subject. This failed our core rule and went in the bin every time.
- Lighting inconsistency. Midjourney occasionally produces lighting from a different angle than the source image implies. This kills photorealism. The fix is explicit atmosphere tagging.
- Signage hallucination. Any AI model will invent text if you let it. Any transform involving visible signage needs Firefly in the pipeline to handle text legibility. Midjourney cannot be trusted here.
- Scale distortion. Big-box transforms would shrink the subject or alter proportions. The preservation rule “preserve: parking lot scale” was added specifically to address this.
✅ What Season 2 Changes
We’re adding a community submission pipeline. If your idea gets selected, you’re tagged in the episode reveal. Submit your idea at the link below.
Submit Your Idea for Season 2
Season 2 production starts in April. We’re opening the submission queue now. Submit a Remodel Idea →