Open editor →

Subject anywhere

Your product photo, placed in any scene — with the label checked.

The coffee bag moved to an alpine terrace with its headline label perfect — and the checker still caught the microprint · 2026-11-17 · Reported saved run: $0.0906. google/gemini-3.8-flash · bria/rmbg-3 · nano-banana-edit · z-ai/glm-5.3-flash

Gemini writes a preservation spec of the product first, including every printed word on it. BRIA RMBG-3.0 cuts the subject out on transparency, so it is lifted rather than repainted. Nano Banana Edit builds the new scene around the untouched cut-out. A second Gemini, blind to the original, re-reads the placed subject, and a GLM checker rules whether the label and the subject survived the move.

Saved result

Cut out · BRIA RMBG-3.0 · reported $0.018 — Subject anywhere
Cut out · BRIA RMBG-3.0 · reported $0.018 · Download
Placed on the alpine terrace · Nano Banana Edit · reported $0.039 — Subject anywhere
Placed on the alpine terrace · Nano Banana Edit · reported $0.039 · Download
Preservation spec · Gemini reads the original · reported $0.0048
SUBJECT: Matte black North Peak coffee bag LABEL TEXT: NORTH PEAK DARK ROAST 250g SMALL BATCH • WHOLE BEAN • RICH & BOLD ROASTED IN THE MOUNTAINS • 100% ARABICA COLOURS: matte black, warm cream SHAPE: Tall upright rectangular standing pouch with side gussets, flat base, and sealed top edge
Preservation spec · Gemini reads the original · reported $0.0048 · Download
Blind re-read · Gemini transcribes the placed subject · reported $0.0288
LABEL TEXT SEEN: NORTH PEAK DARK ROAST 250g SMALL BATCH • WHOLE BEAN • RICH & BOLD ROASTED IN THE MOUNTAINS • 100% ARABICA SUBJECT SEEN: Coffee bag COLOURS SEEN: Black, cream CONDITION: Lower label text is warped and misspelled with garbled lettering in "MOUNTAINS" and "ARABICA"
Blind re-read · Gemini transcribes the placed subject · reported $0.0288 · Download
Preservation verdict · GLM compares spec and re-read
LABEL: PRESERVED SUBJECT: CHANGED — condition reports warped, garbled lettering in "MOUNTAINS" and "ARABICA," contradicting the intact spec text VERDICT: NEEDS WORK — label distortion/regeneration on the lower text lines must be fixed before the placement is acceptable.
Preservation verdict · GLM compares spec and re-read · Download

References that went in

Product photo
Product photo

Upload a product photo, name a scene, and the product is moved without being repainted. Gemini first writes a preservation spec of the original: subject, colours, silhouette, and every printed word transcribed line by line. BRIA RMBG-3.0 — new on NanoGPT today, a matting model built to hold product text, logos and fine hair — cuts the subject out on transparency, so its pixels are lifted rather than redrawn. Nano Banana Edit then builds the new scene around the untouched cut-out under an instruction not to alter it. A second Gemini, blind to the original, re-reads the placed subject, and a GLM checker holds the two reads against each other and rules on the label and the verdict. In the saved run the bag travels from a kitchen counter to a golden-hour alpine cafe terrace and the placement is beautiful: headline label NORTH PEAK / DARK ROAST / 250g letter-perfect at full resolution. The checker still ruled NEEDS WORK — the blind re-read's condition line flagged the bottom microprint ('ROASTED IN THE MOUNTAINS • 100% ARABICA') as warped, and at full size that line is indeed the softest on the bag. That catch is the point of the flow: placement tools usually ask whether a composite looks good; this one asks whether it is still the same product, and prints the answer. Flaws and limits: the placement is a generative edit, so 'untouched' is an instruction the audit verifies, not a pixel guarantee — a true alpha-composite step would make it exact. The audit compares two vision transcriptions, so microprint near reader resolution can be flagged on a soft-but-faithful rendering, as arguably happened here. The cut-out keeps the subject's original lighting; the editor harmonises the scene around it rather than relighting the subject. One subject per run. The demo photo was itself painted by Muse Image in a separate $0.011 run because the flow takes any upload — swap in a real product photo and it runs as saved.

Make it yours

  1. Open the graph — save first if the canvas already has your work. Undo brings yours back.
  2. Swap Product photo for your own product shot and rewrite New scene to wherever it should go. Leave Placement brief and the checker's system prompt in place — they carry the do-not-alter instruction and the verdict shape. Keep Cut out on bria/rmbg-3.
  3. Check the editor estimate, then run with your NanoGPT key. Five model calls: the saved run reported $0.0906 in total (spec $0.0048, RMBG-3.0 cut-out $0.018, Nano Banana placement $0.039, blind re-read $0.0288, GLM checker $0.00). Prices and results vary.
  4. Compare the original, the cut-out and the placed composite, then read the Preservation verdict. PRESERVED means every spec line was re-read exactly; DRIFTED quotes the lines that differ. In the saved run the headline label was letter-perfect and the checker still flagged the microprint line — that strictness is the feature, not a failure of the demo.

Download workflow · Original sampled graph · Explore MCP tools · awesome-noodles