Create a document. preset: 1080p, 4k, square, instagram_post, story, youtube_thumbnail, a4_300dpi, a4_150dpi, web_banner, og_image; or width/height; or from_image= to start from a photo (canvas = photo size). path=<file.json> saves it (autosave after every edit). background can be 'transparent'.

Open a pedit .json document, or a photo (becomes a new unsaved document of its size).

Document state: size, layers (top first) with ids, boxes, effects, masks. detail='full' adds all params.

Save the document (.json). With path: save as.

Document name, background color ('transparent' allowed), dpi.

Undo the last edit.

Redo the last undone edit.

Image layer from a file (path) or an imported asset. fit: contain, cover, width, height, none (ignored if scale given). Without x/y it is centered. above= to insert right above it.

Text layer. font: sans, serif, mono, display, handwriting, or a system font name/file. max_width>0 wraps lines. Without x/y it is centered; then align_layer to place it.

Vector shape: rect (radius = rounded corners), ellipse, line, polygon (points relative to x,y). Colors accept #rrggbbaa for transparency.

Full-canvas fill: solid, linear gradient (angle 90 = left->right, 180 = top->bottom) or radial (center cx, cy as fractions). index=0 puts it at the bottom (a background).

Adjustment layer on top: its filters change everything below it, non-destructively. One filter (type + params) or a chain effects=[{type, params}, ...]. See list_effects.

Brush stroke: polyline [[x,y],...] in document px on a paint layer (created if missing). erase=true erases previous strokes of that layer. Clone stamp: source=[x,y] is where to copy from for the first point (copies everything below the layer); heal=true = healing brush (matches the light and color of the destination).

Layer name, visibility, lock, opacity 0..1, clip (clipping mask onto the layer below), blend mode: normal, dissolve, darken, multiply, color_burn, linear_burn, darker_color, lighten, screen, color_dodge, add (linear dodge), lighter_color, overlay, soft_light, hard_light, vivid_light, linear_light, pin_light, hard_mix, difference, exclusion, subtract, divide, hue, saturation, color, luminosity; groups also pass_through.

Move/scale/rotate (degrees, clockwise)/flip a layer. width or height = target size on the canvas (keeps proportions). crop=[x,y,w,h] in SOURCE pixels of an image layer ([] removes it).

Align to the canvas: h left/center/right, v top/middle/bottom (null = leave), margin in px.

Change a text layer's text or style.

Change a shape layer (fill/stroke 'none' removes them).

Change a fill/gradient layer.

Reorder within its list: to=up/down/top/bottom, or index (0 = bottom). into= moves it inside a group; into="root" takes it out of its group.

Put layers (same level) into a new group. Groups have their own opacity, blend (default pass_through), mask and effects; an adjustment inside a non pass-through group affects only the group.

Dissolve a group: its layers go back in its place.

Copy a layer (placed right above it).

Delete a layer (undoable).

Turn a text/shape/paint layer (with its effects and mask) into pixels.

Merge a layer into the one below.

All visible layers into one image layer.

Effect catalogue. Without name: names and one line each. With name: its parameters.

Filter or layer style on one layer (filters before the mask, styles after: shadow, stroke, glow).

Change an effect's params or switch it off/on. index = position in the layer's effects.

Remove an effect from a layer.

Reorder effects (order matters: denoise before sharpen).

Ready-made looks for apply_look.

Apply a look: as an adjustment layer on top (default) or as effects on one layer.

Layer mask (white = visible). kind rect/ellipse with box=[x,y,w,h] in document px, image with path (grayscale or alpha file, aligned to the layer content), or paint (to paint with mask_stroke; hide_all=true starts fully hidden). feather = soft edge px.

Paint on a layer's mask: mode hide or reveal, polyline [[x,y],...] in document px. Works on any mask (refine an AI cut-out); a layer without a mask gets a paint mask.

Invert, soften, weaken or move an existing mask.

Remove a layer's mask.

Crop the canvas to a box (document px). Layers are kept whole: undoable without loss.

Change canvas size without scaling content. anchor: center, top, bottom_left, ...

Generative expand: add px on each side and fill the new borders (a new top layer masked to them). No prompt = classic fill (fine for sky, walls, plain backgrounds); with prompt or an AI service = generative.

Scale the whole document (canvas and content), proportions kept: pass one side.

LOOK at the result: the real rendered image. box=[x,y,w,h] zooms on a detail (document px). layer= shows that layer alone. grid=true overlays a 10% grid labelled in document px.

Numbers about the image (or one layer, or a box): brightness, contrast, clipped shadows/ highlights, saturation, color cast, sharpness, dominant colors. Measure before correcting.

Check the document: text off-canvas or with low contrast, layers outside, blown highlights, upscaled images... each warning names the tool that fixes it. Not done while it warns.

What AI is available here: background removal, inpaint backend, ComfyUI (generation, upscale), online services with a key (online.ready: service -> tasks) and the default one.

Online AI services (OpenAI, Gemini, Stability, OpenRouter, Replicate, Anthropic, Ollama, custom): which are ready, for which tasks (generate, edit, inpaint, describe, upscale, cutout), with which models. With service: set its models {task: model}, url, or default=true (used by backend="auto"). API keys are pasted by the user in the editor (AI models) or set as env vars, never here.

AI cut-out as a layer MASK (the photo stays intact). keep=background inverts it. Models: isnet-general-use, u2net, u2net_human_seg (people), birefnet-general (best, slow). backend=stability uses the online service instead of the local model.

Bounding box [x,y,w,h] of the main subject of an image layer, in document px.

Remove an object / fill an area: box=[x,y,w,h] document px, mask_path, or points=[[x,y],...] + size (a brush stroke: spot healing). No prompt = remove (classic fill). With prompt = generative fill (ComfyUI or an online service: backend=openai, gemini, stability, ...). The result is a patch on a NEW layer with a soft mask: move, fade or delete it like any layer.

Text-to-image with the local ComfyUI or an online service (backend=openai, gemini, stability, openrouter, replicate), placed on a new layer (fit cover/contain).

Image-to-image: redo an image layer (or a box of the picture, or all of it) following the prompt ("make it winter", "watercolor style"). strength 0..1: low keeps it, high reinvents it. The result is a NEW layer above, same place and mask; the original stays below.

Image-to-text with a vision model (online with a key, or Ollama): describes the picture, a layer or a box, or answers question about it. Does not change the document.

More real pixels for an image layer (ComfyUI model, or backend=stability online, else Lanczos+sharpen). Its size on the canvas does not change.

Measure and correct levels, white balance, vibrance, contrast. Adds normal, editable filters (an adjustment layer on top, or effects on layer=).

Make a layer's colors and tones match another layer (e.g. a pasted subject vs the background).

Point pedit at a ComfyUI and choose its checkpoint/upscaler. Without args: current settings.

Write the image file (.png .jpg .webp .avif .tif). ONLY when the user asks for it. transparent=true keeps alpha (png/webp). box exports a region.

Many edits in ONE call and ONE undo step, all or nothing: [{"tool": "set_transform", "args": {...}}, ...]. look=true returns the preview too.

Open the web editor on THIS document (same in-memory document, live updates).

Stop the web editor started with open_ui.

Log what is broken/clumsy/missing (kind: bug, scomodo, sbagliato, manca, lento), then go on.