Input: dish photos, menu text, and prices
Output: one consistent menu row
Aspect: 1:1 and 4:3
Build log
1. Set the rule
I opened the menu photos, recipe steps, and one hero item, locked the real dish shape, portion, and ingredients, and wrote down what could not change.
2. Run one minimum version
I made one small one consistent menu row first. The menu text warped, so I locked the type and used a flat surface.
3. Build the focused variant
I reused the same reference and changed only one menu item and one readable price block.
4. Review and export
I checked dish shape, portion, texture, and no invented ingredients at full size, then exported the final 1:1, 3:4, and 9:16.
Complete execution record
Task boundary and source gate
Treat AI restaurant menu image as a deliverable, not a definition. The job receives dish photos, menu text, and prices and owes one consistent menu row in 1:1 and 4:3. Write the rejection reason first, then turn it into the first hard constraint.
Require the plated reference, portion size, ingredient list, garnish, table surface, kitchen constraint, and the approved menu wording. Do not invent hygiene claims, health effects, or a preparation step the kitchen does not follow. Recipe showmanship must not replace safe handling.
The imgmov route
Upload the plated dish, hero ingredient, and recipe order. Use cheap generation to prove steam, pour, cut, plating, or texture. Menu facts, price, allergen wording, and safety notes go into the external editor.
FDA separates safe handling into clean, separate, cook, and chill, and says color and texture are unreliable safety indicators. A recipe clip may show process, but safety wording must come from the approved source.
Lock the variables that cannot drift
Lock portion size, hero ingredient color, garnish, plate shape, sauce state, table surface, and any food-safety step. A clip cannot increase the portion.
Put the locked fields at the top of the prompt, not at the end: The practical move is to lock the real dish shape, portion, and ingredients and change only one menu item and one readable price block. Save that prompt with the reference, ratio, and model so the next run starts from a decision instead of a guess.
Step-by-step build path
1) Set the rule: I opened the menu photos, recipe steps, and one hero item, locked the real dish shape, portion, and ingredients, and wrote down what could not change. Leave one checkable artifact from this step; do not start the next until it exists. 2) Run one minimum version: I made one small one consistent menu row first. The menu text warped, so I locked the type and used a flat surface. Leave one checkable artifact from this step; do not start the next until it exists. 3) Build the focused variant: I reused the same reference and changed only one menu item and one readable price block. Leave one checkable artifact from this step; do not start the next until it exists. 4) Review and export: I checked dish shape, portion, texture, and no invented ingredients at full size, then exported the final 1:1, 3:4, and 9:16. Leave one checkable artifact from this step; do not start the next until it exists.
Change one named variable per rerun: prompt, reference, camera, duration, model, ratio, or export crop. If two variables change together, a better output cannot be reused because nobody knows which fix worked.
Review gates and evidence
The hero ingredient must be identifiable, texture must not become plastic, hands must not melt, and menu captions must match what is actually served.
Keep source, approved wording, rejected version, correction, credit cost, and final crop next to the asset. The record exists so the next reviewer can reproduce the decision without a chat thread.
Failure diagnosis and retry ladder
If texture becomes plastic, reduce motion and use a macro still. If sauce flows forever, shorten the pour. If garnish changes, return to the plated reference.
Image retries have a ladder. Start with Agnes Image 2.5 for the cheap proof. Use SDXL Inpainting at 3 credits when only a masked region is wrong. Use Seedream 5.0 Lite or Nano-Banana 2 at 5 credits for a more complex rebuild. GPT Image 2 starts at 2 credits for Standard 1K and rises to 80 by quality and resolution. Name the failure before changing rungs.
Versioning and handoff
Menu boards, delivery apps, and short video get separate safe-area checks because price and spicy-level text crop differently.
The folder is boring on purpose: source, approved reference, generation settings, caption file, platform cuts, QA screenshots, and one correction line. A reusable asset is boring in the right way.
Why this is not a generic answer
Editing is not detached from production. Object detection, masks, reference images, version history, Canvas dependencies, and the final generation set stay together. A standalone generator can invent a scene; it will not preserve the source asset, the rejected versions, and the reason the approved crop won.
Do not invent hygiene claims, health effects, or a preparation step the kitchen does not follow. Recipe showmanship must not replace safe handling. If the request only asks what ai restaurant menu image means, a search page is faster. This page is useful when someone must deliver one consistent menu row under real constraints.
Frequently asked questions
What is the one rule I keep repeating?
The practical move is to lock the real dish shape, portion, and ingredients and change only one menu item and one readable price block.
The workflow is specific about its stopping point: it starts with dish photos, menu text, and prices and stops at one consistent menu row. In imgmov, upload the source, lock it as a reference, set 1:1 and 4:3, run one proof, and save the passing settings. If a reviewer cannot tell which version is approved, the workflow has failed even when the media looks good.
What do I check before export?
Portion, ingredients, texture, text, and final crop.
Image retries have a ladder. Start with Agnes Image 2.5 for the cheap proof. Use SDXL Inpainting at 3 credits when only a masked region is wrong. Use Seedream 5.0 Lite or Nano-Banana 2 at 5 credits for a more complex rebuild. GPT Image 2 starts at 2 credits for Standard 1K and rises to 80 by quality and resolution. Name the failure before changing rungs.
What makes this different from a generic generator?
I start from dish photos, menu text, and prices and keep one consistent menu row consistent, instead of inventing a new scene each run.
The check is not “does it look AI-nice?” It is: The hero ingredient must be identifiable, texture must not become plastic, hands must not melt, and menu captions must match what is actually served. Then the source, approved copy, rejected version, correction, and final crop stay in the same handoff folder.
What should I do if the first render drifts?
Cut the motion, lock the reference, and rerun one small version before batching.
If texture becomes plastic, reduce motion and use a macro still. If sauce flows forever, shorten the pour. If garnish changes, return to the plated reference. Change one named variable, keep the old version, and record the credit cost. If the same failure repeats, fix the reference or scope rather than asking the prompt for forgiveness.