Frontier generators like GPT Image 2 and Nano Banana still accumulate noise when you feed each edit back into the model.
Pointlight separates local changes from a dedicated HD redraw pass — clean the finished frame without regenerating the whole room.
Drag to compare. Same camera, same furniture placement — only clarity changes.
How the noise and artifacts form
Pass one from a modern image model often looks sharp. The failure mode shows up in the multi-turn loop:
you take the output, ask for another change, and feed that result back in. Each pass re-samples the image.
Tiny mismatches — soft edges, color shift, texture grit — stack until the living room looks “AI soft” even when the layout is still roughly right.
1. Error accumulation
Diffusion-style generators are sequential. Prediction error from earlier steps compounds into later ones — formalized as modular vs cumulative error in diffusion architectures (ICLR 2024).
2. Iterative degradation
On Nano Banana Pro, the Banana100 study (CVPR 2026 workshop) shows visible noise, tint, and texture loss within about 5–10 edit steps — and the noise persists even when you prompt “denoise.”
3. Global re-edit drift
Asking GPT Image 2 / Nano Banana 2 to “fix the whole image again” reopens composition. Sofas move, windows shimmer, materials re-roll — not just sharpness.
4. Inpaint seam artifacts
Region edits can leave local patches that don’t match surrounding grain. Stack a few of those and the frame looks noisy even before a full regen (GradPaint discusses harmonization failures in diffusion inpainting).
So this is not “your prompt was weak.” It is a known failure mode of multi-turn generative editing.
Frontier models still win on first-pass beauty; they do not automatically win the tenth client revision.
What Pointlight does differently
We do not claim a secret model that erases physics or makes GPT Image 2 / Nano Banana obsolete.
Pointlight’s answer is a revision-first product loop built for interiors:
Lock a base from geometry, photo, or an accepted render — structure first.
Edit locally for client notes (sofa, rug, wall) so you are not re-rolling the whole room every time.
HD redraw last — a dedicated cleanup pass aimed at softness and noise on the finished frame, without treating “regenerate everything” as the fix.
That last step is the product feature you see in the before/after above: same camera, same furniture placement, higher clarity.
It is engineering for the revision loop — not a “world’s only AI” claim.
How to run HD redraw in Pointlight
Once the design is locked and only clarity is wrong, stay in the app — don’t start a new full generation. Three clicks:
Step 1
Open HD redraw
In the top modes bar, choose HD redraw (next to Lighting / Room staging 3D).
Step 2
Set Source Render
Use the finished image you just edited — the one that looks soft or noisy after local changes.
Step 3
Pick Standard or Advanced
Run from the green buttons at the bottom of the left panel. Start with Standard.
HD Redraw (Standard) 1 credit
Everyday cleanup after a few local edits — first pass for most client rounds.
HD Redraw (Advanced) 2 credits
Stronger redraw when Standard still leaves softness or noise on materials and edges.
Tip: run HD redraw last. If the chair is in the wrong place, fix that with a local edit first.
FAQ
Why does quality drop after several AI edits?
Each generative pass re-samples the image. Small artifacts compound into visible noise, tint, or texture loss. Banana100 shows this on Nano Banana Pro within roughly 5–10 turns; diffusion research describes cumulative error across sequential steps.
Do GPT Image 2 and Nano Banana still have this problem?
On multi-turn “edit the output again” loops, yes. Banana100 documents persistent noise on Nano Banana Pro even with denoise prompts. GPT Image 2 is strong on single passes, but global re-edits reopen the same accumulation risk. Pointlight splits local edits from a final HD redraw cleanup.
Is HD redraw just upscaling?
No. Upscaling enlarges. HD redraw is a dedicated cleanup pass on the finished frame for clarity while aiming to keep composition and placement intact.
Does this work with SketchUp screenshots?
Yes. Viewport screenshots, room photos, and prior AI passes all work as a base. SketchUp is an input option — not a requirement.
Standard vs Advanced — which should I pick?
Start with Standard (1 credit). If materials or edges still look soft, run Advanced (2 credits). Don’t jump to Advanced to fix a wrong layout — edit locally first.
When should I not use HD redraw?
If the layout itself is wrong, fix placement with a local edit first. HD redraw is for when design is locked and the image needs clean detail.
