imageai-cli

name

imageai-cli

description

Use when generating, editing, batching, or scripting images, videos, or page layouts with the ImageAI CLI (`python main.py ...`) from inside the ImageAI repo, across any provider — Google Gemini / Nano Banana, OpenAI (gpt-image-2/1.5/1, DALL·E), Stability AI, or local Stable Diffusion. Covers model selection, sizing/aspect, references & masks, streaming, Batch API, lyrics-to-prompts, gcloud auth, key management, video generation (Gemini Omni + Veo: refs, extend, frame interpolation, conversational refine, edit-your-own-video, --json output contract), and the publication layout engine (--layout-design/--layout-fill/--layout-export for comics, magazines, PDFs). Trigger whenever the user wants to make or edit an image, video, or layout (or set up/test a provider key) from the command line in this working directory.

imageai-cli

The cheat-sheet for driving everything the ImageAI CLI can do from python main.py. Four image providers, two video providers, the publication layout engine, the lyrics-to-prompts pipeline, the OpenAI Batch API, and key/auth management — all from one entry point.

GUI is the default when no action flag is given. Add --gui to force it, or pass -p/--prompt, --video, -t/--test, -s/--set-key, --lyrics-to-prompts, or any --layout-* action to stay in the CLI.

Full documentation (agent recipes, output contracts, exit codes): Docs/ImageAI-CLI-Guide.md.

Pick a provider first

Provider	--provider	Auth	Best for
Google Gemini	google (default)	API key or gcloud ADC	Nano Banana family; fast, high-quality, up to 4K (NBP)
OpenAI	openai	API key	gpt-image-2 "thinking" model, DALL·E 3
Stability AI	stability	API key (hosted)	SDXL / SD hosted endpoints
Local SD	local_sd	none (local GPU/CPU)	offline generation, no per-image cost

Default provider is google, default model gemini-2.5-flash-image.

Universal invocation

# Minimal: default provider/model, auto-named output in the images dir
python main.py -p "a red fox in snow"

# Explicit provider + model + output path
python main.py --provider <prov> -m <model> -p "your prompt" -o out.png

# N images in one call
python main.py --provider openai -m gpt-image-2 -n 4 -p "logo ideas" -o logo.png

-o/--out is optional — without it, images auto-save with a sanitized, prompt-derived filename under the platform images dir. Each image gets a .json metadata sidecar.

---

Google Gemini (Nano Banana)

Model	Alias	Max output
gemini-3-pro-image-preview	Nano Banana Pro	4K
gemini-3.1-flash-image-preview	Nano Banana 2	2K
gemini-2.5-flash-image (default)	Nano Banana	1024px

# Default Nano Banana
python main.py -p "watercolor harbor at dawn" -o harbor.png

# Nano Banana Pro at 16:9 (provider maps --size to an aspect ratio)
python main.py -m gemini-3-pro-image-preview --size 1920x1080 \
    -p "cinematic desert highway" -o road.png

# gcloud Application Default Credentials instead of an API key
python main.py --auth-mode gcloud -p "..." -o out.png

• Aspect ratio, not pixels in the prompt. Set sizing via --size; the provider converts to an aspect ratio (image_config). Never bake dimensions like "(1024x768)" into prompt text — Gemini renders them as literal text.
• Supported aspect ratios: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9.
• Scaling: NB caps at 1024px; the app scales proportionally (max edge 1024) then upscales. NB2 → 2K native, NBP → 4K native.
• gcloud auth (--auth-mode gcloud) works for the Gemini image models with no API key; run gcloud auth application-default login first.
• Operations: generate, edit, compose (pass --reference for edit/compose).

OpenAI

Model	Notes
gpt-image-2	"Thinking" model — --quality drives reasoning. Best quality.
gpt-image-1.5	Latest non-thinking; supports --output-format.
gpt-image-1	Prior gen.
gpt-image-1-mini	Fast/cheap.
dall-e-3	`--quality standard
dall-e-2	Legacy.

# gpt-image-2, explicit reasoning
python main.py --provider openai -m gpt-image-2 --quality high \
    -p "complex composition with legible text" -o gen.png

# DALL·E 3 in HD
python main.py --provider openai -m dall-e-3 --quality hd \
    -p "isometric city" -o city.png

See the gpt-image-2 deep-dive below for custom sizes, masks, streaming, and batch — it has the most surface area.

Stability AI (hosted)

Model	Label
stable-diffusion-xl-1024-v1-0 (default)	SDXL 1.0
stable-diffusion-xl-beta-v2-2-2	SDXL Beta
stable-diffusion-512-v2-1	SD 2.1
stable-diffusion-v1-6	SD 1.6

python main.py --provider stability -m stable-diffusion-xl-1024-v1-0 \
    -p "fantasy castle, concept art" -o castle.png

Needs a Stability API key (--provider stability -s to store one; -t to test).

Local Stable Diffusion (offline)

Model	Label
stabilityai/stable-diffusion-xl-base-1.0	SDXL Base 1.0
segmind/SSD-1B	SSD-1B (Fast SDXL)
stabilityai/stable-diffusion-2-1 (default)	SD 2.1
runwayml/stable-diffusion-v1-5	SD 1.5
CompVis/stable-diffusion-v1-4	SD 1.4

python main.py --provider local_sd -m segmind/SSD-1B \
    -p "studio portrait, soft light" -o portrait.png

• No API key. First run downloads weights from Hugging Face (slow, large).
• Uses GPU (CUDA / Apple MPS) when available, falls back to CPU (slow).
• Requires the local-SD extras (torch, diffusers) to be installed; if they're missing, tell the user what to install rather than installing it yourself.

---

gpt-image-2 deep-dive

ImageAI ships first-class support for OpenAI's gpt-image-2 (released 2026-04-21).

# Custom size (edges multiples of 16, max 3840, aspect ≤3:1, pixels 655K-8.3M)
python main.py --provider openai -m gpt-image-2 \
    --custom-size 2048x1152 -p "ultrawide wallpaper" -o gen.png

# Multi-reference compose (up to 10 images) → /v1/images/edits
python main.py --provider openai -m gpt-image-2 \
    --reference ref/a.png --reference ref/b.png \
    -p "combine these styles" -o composed.png

# Mask inpainting (transparent pixels in mask = edit zone)
python main.py --provider openai -m gpt-image-2 \
    --reference base.png --mask mask.png \
    -p "replace the sky with stormy clouds" -o edited.png

# Streaming partials (writes gen.p0.png, gen.p1.png, then gen.png)
python main.py --provider openai -m gpt-image-2 \
    --stream-partials --quality high \
    -p "intricate technical diagram" -o gen.png

# JPEG output with compression
python main.py --provider openai -m gpt-image-2 \
    --output-format jpeg --output-compression 85 \
    -p "photorealistic landscape" -o landscape.jpg

# Permissive moderation
python main.py --provider openai -m gpt-image-2 --moderation low -p "..." -o gen.png

Cost estimator (1024×1024)

Quality	Per image	At n=10
low	~$0.006	~$0.06
medium	~$0.053	~$0.53
high	~$0.211	~$2.11
Batch	50% off all of the above

Costs scale with output tokens; custom sizes bigger than 1024² cost more.

Snapshot pinning

from core.constants import GPT_IMAGE_2_SNAPSHOT  # "gpt-image-2-2026-04-21"

Pass the snapshot ID as -m when you need bit-for-bit reruns across releases.

---

Batch API (OpenAI)

# Submit (50% discount, up to 24h turnaround) — prints a job ID
python main.py --provider openai -m gpt-image-2 --batch -p "your prompt" -o gen.png
# → Submitted batch job: batch_abc123...

python main.py --provider openai --batch-status batch_abc123
python main.py --provider openai --batch-fetch  batch_abc123

Jobs are tracked in ~/.imageai/batch_jobs.json (also surfaced in the GUI's Batch Jobs tab).

Lyrics → prompts (LLM pipeline)

Turns a lyrics file into a set of image prompts using an LLM (not an image model).

python main.py --lyrics-to-prompts song.txt \
    --lyrics-model gpt-4o \
    --lyrics-style cinematic \
    --lyrics-temperature 0.7 \
    --lyrics-output prompts.json

--lyrics-model accepts any LiteLLM id (gpt-4o, gemini/gemini-2.0-flash-exp, claude-sonnet-4-6, …). Feed the resulting prompts back into image generation.

Video generation (--video)

Two providers: --video-provider omni (Gemini Omni, conversational, audio baked in) and veo (default). Reuses -p/--prompt and -o/--out.

# Veo text-to-video (default provider)
python main.py --video -p "a fox running through snow" -o fox.mp4

# Omni text-to-video (16:9 default; audio auto-generated)
python main.py --video --video-provider omni -p "a marble run, smooth shot" -o marble.mp4

# Subject references (up to 3 images for either provider)
python main.py --video --video-provider omni -p "the cat bats the yarn ball" \
    --ref-image cat.png --ref-image yarn.png -o cat.mp4

# Extend an existing clip (Veo only; extensions render at 720p)
python main.py --video --extend fox.mp4 -p "the fox leaps over a log" -o fox2.mp4

# Frame-to-frame interpolation (Veo only: start ref + end frame)
python main.py --video -p "sunrise timelapse" -o sunrise.mp4 \
    --ref-image start.png --last-frame end.png

# Conversationally refine the previous clip (Omni; id = operation_id in the sidecar)
python main.py --video --video-provider omni --refine-from int_abc123 \
    -p "make the violin invisible" -o v2.mp4

# Edit your own footage (Omni; uploads via the Files API first)
python main.py --video --video-provider omni --edit-video input.mp4 \
    -p "make the mirror ripple like liquid" -o edited.mp4

# Large/720p clips: ask for URI delivery (Omni only)
python main.py --video --video-provider omni --delivery uri -p "city timelapse" -o city.mp4

# Machine-readable result (exactly one JSON object on stdout; logs on stderr)
python main.py --video -p "a fox in snow" -o fox.mp4 --json

Provider capability split

Capability	Omni	Veo (default)
--ref-image (≤3)	✅	✅
--last-frame interpolation	❌	✅
--extend existing clip	❌	✅ (720p)
--refine-from / --edit-video / --delivery uri	✅	❌
Audio in output	✅ always	✅ (Veo 3)
Aspect ratios	16:9, 9:16	16:9, 9:16, 1:1
--auth-mode gcloud	❌ (API key only)	✅

Clip length is model-fixed (~8 s) — no --duration knob; chain --extend (Veo) for longer sequences. --video-model overrides the resolved default model. Wrong provider/flag pairings fail fast with exit code 2.

Omni prompt superpowers (in the prompt text, not flags)

• Image roles: <FIRST_FRAME> (start frame) vs <IMAGE_REF_N> (subject refs, N=0,1,2); declare with [# Sources <FIRST_FRAME>@Image1].
• Timing: natural language ("after 3 seconds, ...") or timecodes ([0-3s] a woman enters, "every 2s cut to a new angle").
• Audio direction: describe music/sound design in the prompt — Omni soundtracks every clip by default.
• On-screen text: Omni renders readable/animated text ("a neon sign that reads OPEN").
• Aspect ratio goes in --aspect (16:9 or 9:16), never in the prompt.

Omni does not support: video extension (--extend is Veo-only), video references > 3s, negative-prompt configs, temperature/system instructions.

Custom styles (--style*)

Derive a reusable named style from your own reference images (vision-LLM analysis), then apply it text-only or with attached exemplar images across image generation, video, and layout fill. Full guide: Docs/CustomStyles.md.

# Create from a folder/files/globs; prints the new style id on stdout
python main.py --style-create "Watercolor" --style-images ./refs/watercolor/
python main.py --style-create "Neon Noir" --style-images shot1.png "refs/noir_*.jpg" \
    --style-llm-provider openai --style-llm-model gpt-4o

# Manage
python main.py --style-list
python main.py --style-show "Watercolor"
python main.py --style-delete "Watercolor"

# Apply (image gen, video, layout fill)
python main.py -p "a lighthouse at dusk" --style "Watercolor" -o lighthouse.png
python main.py -p "a lighthouse at dusk" --style "Watercolor" --style-smart -o lighthouse.png
python main.py --video -p "waves rolling in" --style "Watercolor" -o waves.mp4
python main.py --layout-fill comic.json --style "Neon Noir"

# Share
python main.py --style-export "Watercolor" -o watercolor.zip
python main.py --style-import watercolor.zip

• Google/OpenAI gpt-image-* image generation additionally attaches the style's starred exemplar images as references (after any of your own — yours always win, on limit or when you already have an active reference). Stability AI, Local SD, video, and layout fill are text-only.
• --style-smart (fuse prompt+style via one LLM call, falls back to plain concat on any failure) is image-generation only; ignored elsewhere.
• Unknown --style NAME exits 2 and lists available styles.

Publication layout engine (--layout-*)

Design a page from a text description, fill its image regions, export to PDF/PNG — fully headless.

# Design a layout project (.json) from a description (requires -o)
python main.py --layout-design "3-panel comic, action scene" -o comic.json \
    --content-kind comic --page-size "US Comic" --orientation portrait --dpi 300

# Generate images for every prompted region in the project
python main.py --layout-fill comic.json

# Render to PDF or PNG (the -o extension picks the format)
python main.py --layout-export comic.json -o comic.pdf

• --content-kind: children | comic | comic_strip | magazine | newspaper | scientific | custom (default custom).
• --page-size (Letter, A4, A5, Tabloid, US Comic, …), --orientation (portrait/landscape), --dpi (default 300).
• --layout-design uses a text LLM; override with --layout-llm-provider (openai | anthropic | google | ollama | lmstudio) and --layout-llm-model.
• --layout-fill uses the normal image --provider/-m selection.

The project JSON is hand-editable (schema 2.0)

For fine placement fixes, edit the project file directly and re-export — no GUI needed. Layout: pages[N].regions[] and pages[N].overlays[].

• Page pixels = inches × dpi (e.g. 6.625×10.25 in @ 300 → 1988×3075).
• Region bbox is [x, y, w, h] in page pixels.
• Regions carry image_ref (absolute path to the art; fills land in the platform images dir, e.g. ~/.config/ImageAI/images/), prompt, z.
• Overlays (caption/speech/sfx) place by anchor [x, y] + anchor_mode: center centers the whole text block on the anchor (so "move the title" = set the anchor to the midpoint of the space it should occupy); topleft pins the block's corner. Text look lives in text_style; box look and max_width_px in style.

Region images are cover-cropped — match aspect or lose art

--layout-export fills each region by scaling + center-cropping the art to the region's aspect. Art whose aspect ≠ bbox aspect silently loses its edges — 16:9 art in a ~6:5 comic panel loses ~35% of its width, amputating any off-center subject (symptom: "this panel looks cropped weird").

• Prevent: before filling, compute each region's w/h and generate its art at the closest supported ratio (Google: 1:1, 3:2, 4:3, 5:4, 16:9, …).
• Prevent (Nano Banana): alternatively, tell the model the target resolution in the prompt as a composition instruction — e.g. "compose the scene to fill an 858×720 area" — and it confines the content to that area, so the renderer's crop can't cut anything that matters. This is a where to put content instruction, distinct from the sizing footgun below.
• Repair without regenerating: pre-crop the source to exactly the bbox aspect — full source height, width = height × (w/h), slide the x-window to protect the subject — save as a new file and point image_ref at it. The renderer's center-crop then becomes a no-op. Never overwrite the original art.

Verify visually, headless

After any placement change: --layout-export proj.json -o page.png, crop the area of interest with PIL, and look at it. Edit JSON → export → inspect is a cheap loop; don't trust coordinates alone.

.iaibundle is GUI-only

The CLI exports PNG/PDF only. A project's .iaibundle is saved from the GUI and goes stale when the CLI edits the project — re-save it there if it must stay in sync.

Agent output contract

• --json (video): stdout carries exactly one JSON object (status, output_path, provider, model, aspect_ratio, operation_id, error); all logs/progress go to stderr. Parse stdout only.
• Exit codes (video): 0 success, 1 generation failed, 2 validation / user error (bad flag combo, missing key), 3 unexpected exception.
• Sidecars: every image and video gets a <output>.json sidecar with prompt, resolved model, and ids — durable metadata even without --json.

Keys, testing & auth

# Store a key for the selected provider (interactive)
python main.py --provider openai -s

# Provide a key inline or from a file for a one-off run
python main.py --provider stability -k "$STABILITY_KEY" -p "..." -o out.png
python main.py --provider openai -K /path/to/key.txt -p "..." -o out.png

# Verify the configured key works
python main.py --provider openai -t

# Setup instructions
python main.py --help-api-key

Key resolution order: CLI flag (-k/-K) > stored config > environment. --auth-mode gcloud is Google-only and bypasses API keys via ADC.

---

Full flag reference

Flag	Values	Notes
--provider	google | openai | stability | local_sd	Default google.
-m, --model	provider model id	See per-provider tables.
-p, --prompt	text	Triggers CLI generation.
-o, --out	path	Optional; auto-named if omitted.
-n, --num-images	int	Default 1. gpt-image-2 up to 10.
--size	WxH or preset	Mutex with --custom-size. Google maps to aspect.
--custom-size	WxH	gpt-image-2 only. Edges ×16, max 3840, aspect ≤3:1, pixels 655K–8.3M.
--quality	auto | low | medium | high | standard | hd	gpt-image-2: auto/low/medium/high; dall-e-3: standard/hd.
--output-format	png | jpeg | webp	gpt-image-2 / gpt-image-1.5 only.
--output-compression	0..100	jpeg/webp only.
--moderation	auto | low	gpt-image-2 only; low = permissive.
--reference	path (repeatable ≤10)	Routes to /v1/images/edits.
--mask	PNG path	Alpha mask inpainting; transparent = edit zone.
--stream-partials	flag	gpt-image-2 only; writes out.pN.png.
--batch / --batch-status / --batch-fetch	flag / JOB_ID	OpenAI Batch API.
--lyrics-to-prompts	LYRICS_FILE	Plus --lyrics-model/-temperature/-style/-output.
--video	flag	Video mode; reuses -p/-o.
--video-provider	omni | veo	Default veo.
--video-model	model id	Optional override of the resolved default.
--aspect	e.g. 16:9, 9:16	Validated per video provider.
--ref-image	path (repeatable ≤3)	Video reference images.
--last-frame	path	Veo-only end frame (interpolation).
--extend	path	Veo-only: extend this existing clip.
--delivery	inline | uri	Omni-only; uri for large/720p clips.
--refine-from	INTERACTION_ID	Omni-only conversational refine.
--edit-video	path	Omni-only: upload + edit your own footage.
--json	flag	Video: one JSON object on stdout.
--layout-design	DESCRIPTION	New layout project (requires -o).
--layout-fill	PROJECT	Generate images for prompted regions.
--layout-export	PROJECT	Render to PDF/PNG (requires -o).
--content-kind / --page-size / --orientation / --dpi	see layout section	--layout-design options.
--layout-llm-provider / --layout-llm-model	see layout section	Text-LLM override for design.
--style	NAME	Apply a saved style (image gen, --video, --layout-fill).
--style-smart	flag	Smart-merge prompt+style via LLM (image gen only; falls back to plain concat).
--style-create	NAME	Derive a new style (needs --style-images).
--style-images	PATH ... (repeatable)	Files/dirs/globs for --style-create.
--style-llm-provider / --style-llm-model	provider / model id	Vision LLM override for --style-create.
--style-list	flag	List saved styles.
--style-show	NAME	Print one style's full JSON record.
--style-delete	NAME	Delete a style (and its reference images).
--style-export	NAME	Export to a zip (use -o FILE.zip).
--style-import	FILE	Import a style zip.
--auth-mode	api-key | gcloud	Google-only (images + Veo; not Omni).
-k/-K/-s/-t	—	Key inline / from file / store / test.
--help-api-key	flag	API-key setup instructions.
--version	flag	Print version and exit.
--gui	flag	Force the GUI.

Anti-footguns

• Provider/model mismatch. A model id only works with its provider. dall-e-3 needs --provider openai; gemini-* needs --provider google (or default).
• gpt-image-2 has no transparent background, no input_fidelity, no variations endpoint, no style. The provider raises clear errors. For alpha PNG output use gpt-image-1.5/gpt-image-1.
• gpt-image-2 custom-size: the #1 mistake is non-multiple-of-16 edges. Pre-flight validation rejects it; the GUI shows a live red label.
• gpt-image-2 reasoning is just --quality (auto/low/medium/high) — there is no separate reasoning_effort knob. Higher quality = more thinking compute.
• Org Verification gate: gpt-image-2 requires OpenAI Organization Verification. If --provider openai -t returns a verification message, complete it at https://platform.openai.com/settings/organization/general.
• Google sizing: never put pixel dimensions in the prompt text as a size request ("(1024x768)" renders as literal text in the image); use --size. The one sanctioned use of dimensions in a prompt is the Nano Banana fill-area composition instruction in the layout section.
• --quality standard/hd are DALL·E values; --quality auto/low/medium/high are gpt-image-2 values. Don't cross them.
• local_sd first run downloads multi-GB weights and may need GPU extras — warn the user; don't pip install system deps without asking.
• Video flag/provider pairing: --extend/--last-frame need veo; --refine-from/--edit-video/--delivery need omni. Mismatches exit 2.
• In --json mode, parse stdout only — logs and progress are on stderr by design; never grep stdout for progress text.
• Layout art is cover-cropped to the region's bbox aspect. Generate panel art at a matching ratio (or pre-crop and swap image_ref) — a mismatch silently cuts content off the panel edges.
• --layout-design and --layout-export require -o; --layout-fill writes into the project file in place.

GUI entry points

• Generate tab — provider + model pickers; for gpt-image-2 the quality buttons flip to Low | Medium | High | Auto, and output-format / moderation rows appear.
• Resolution selector — "Custom…" opens a W/H dialog with live validation.
• Generate → Submit as Batch Job… — submits via the OpenAI Batch API.
• Batch Jobs tab — lists jobs from ~/.imageai/batch_jobs.json with Check/Download.
• Settings tab — per-provider API keys and auth mode.