Cut the subject out, including the hair and reflections a colour test cannot reach.

Raise resolution and recover detail — waiting on a real upscaling endpoint.

Swap what is behind the subject and keep the light matching.

Take lettering out and rebuild what was behind it — never a watermark.

Paint over an area, say what belongs there, and leave the rest untouched.

Extend a picture past its edges — after checking a crop would not do.

Lay pictures out side by side for free, or merge them into one with a model.

Turn a phone photo of a product into a studio shot — without changing the product.

Change where the light comes from — and the shadows that follow.

Add colour to a black and white photograph — every hue is an inference.

Change the words in a picture and keep the typeface, perspective and surface.

Re-render the same subject from a different camera angle.

Keep the picture and change how it is made.

Render a rough drawing — your layout, real materials.

Smooth blocking artefacts and recover the edges underneath — it does not un-blur a face.

Clean the cast shadows off a product shot — and keep the contact shadow.

One reference in, the same character from several angles out.

A shot list for one subject, and a single contact sheet if you want to see it.

Remove sensor noise without smoothing away real texture.

Change the weather or the season and keep the place.

9:16, 16:9, square or free — cropped at full resolution, in your own browser.

WebP, JPEG or PNG with a quality slider — and the bytes never leave your browser.

Write a prompt for your own photo and open it in the Generator — nothing is uploaded here.

Click a colour, get a PNG with it transparent — chroma keying in your browser.

The colours actually in a picture — and the words a prompt can use for them.

Two local images into one PNG — side by side or split by a divider.

A few canned looks on canvas — curves, not a colourist.

Thirds, safe area and a centre cross over your frame — burned in only if you ask.

Trace a logo or line drawing into real vector paths — in your browser, no upload.

Read the prompt and settings a generator wrote into a file — locally, and never the location.

Drag to look around an equirectangular photo — a viewer, not a generator.

Draw a route on a still and export the overlay — a drawing, not a video step.

Hand it a picture and get a prompt that would produce one like it.

House order for free, a rewrite in Visxy dialect when you want one.

Characters, words, clauses and an honest token range — before a model quietly truncates the end of your prompt.

Translate a prompt without translating its flags, weights or camera notation.

Structure and syntax problems a rule can actually see — duplicate clauses, mixed dialects, negation in a positive prompt.

Templates for the failure modes diffusion models actually have — pick a subject, copy the list.

Which of the seven axes your prompt states, and which the model is deciding for you.

One template with {braces}, every combination expanded in your own browser.

Read, convert and nudge emphasis syntax — the two dialects people mix, normalised.

Remove what a model does not read — and see exactly what came out.

Midjourney flags and :: weights, converted both ways — without rewriting a single word of yours.

A schema form that emits a structured prompt — three families, and nothing filled in by a model.

Which of the four libraries does this belong in — and the words that decided.

Lift negation out of a positive prompt and add the failure set for what it describes.

Raise the resolution of an existing clip. Waiting on a video ingest path and a real upscaler.

Drive a still with the motion of a reference video. Waiting on the video-reference attach product, not on a price.