picture-it（想象一下）

picture-it

Photoshop for AI agents. Composable image operations from the CLI.

Source: https://github.com/geongeorge/picture-it | npm: https://www.npmjs.com/package/picture-it

Prerequisites

picture-it must be installed and configured. Requires Node.js 18+.

# Install (pick one)
npm install -g picture-it
pnpm add -g picture-it
bun install -g picture-it

# Setup
picture-it download-fonts

Credentials

The FAL API key is required for AI operations (generate, edit, remove-bg, upscale). Set it via environment variable or the CLI:

# Option 1: Environment variable (preferred — use platform-managed secrets)
export FAL_KEY=your-key-here

# Option 2: CLI config (stored in ~/.picture-it/config.json with 0600 permissions)
picture-it auth --fal <fal-api-key>

NEVER paste API keys into chat. Always use environment variables or the CLI auth command. Get a FAL key from https://fal.ai.

Note: User images are uploaded to fal.ai for AI processing when using generate, edit, remove-bg, or upscale commands. Local-only commands (crop, grade, grain, vignette, text, compose, template, info) do not transmit data.

Core Concept

Every command takes an image in and outputs an image. Chain them to build anything. The agent calling picture-it IS the planner — there is no AI planner inside the tool.

Before You Generate Anything — Think First

Image generation costs real money ($0.03–$0.15 per FAL call). A 4-pass workflow is $0.10+. Don't burn budget on a vague idea — spend time planning before running any commands.

Step 1: Understand the purpose

Before touching picture-it, get full clarity on what the user wants. Ask yourself:

• What is this image for? (blog header, Instagram ad, YouTube thumbnail, product comparison, poster)
• Who is the audience? (developers, consumers, enterprise buyers)
• What should someone FEEL when they see it? (excitement, trust, urgency, curiosity)
• What's the one message? Every good image communicates exactly one thing.
• Where will it be displayed? This determines size, text sizing, and composition rules.

If any of these are unclear, ask the user before proceeding. A 30-second question saves $0.15 in wasted generation.

Step 2: Plan the composition

Think through at least 3 different approaches before picking one. Consider:

• Can this be done without FAL? Templates and Satori compose are free. A solid gradient + good typography is often enough.
• What's the minimum number of FAL calls? Each call costs money. Plan the fewest passes that achieve the goal.
• Which technique fits? Text-behind-subject for thumbnails, remove-bg + compose for product photos, multi-pass for cinematic scenes.

Present your top 2-3 ideas to the user briefly — one sentence each — and let them pick before generating. Example:

"Here are a few directions: 1. Dramatic product shot — generate a dark stage, edit to place your logo as a glowing 3D object ($0.07) 2. Clean comparison — remove-bg from both products, compose on gradient with text ($0.01) 3. Text-behind-subject — generate an action scene, edit to weave the title behind the subject ($0.07) Which direction, or a mix?"

Step 3: Plan the pipeline

Before running the first command, write out the full pipeline:

1. generate (flux-dev $0.03) — dark stage scene
2. edit (seedream $0.04) — place logo into scene
3. compose (free) — add text overlay
4. grade + vignette (free) — post-process
Total: ~$0.07

This avoids discovering mid-way that you need a different approach and wasting the earlier calls.

Commands Quick Reference

Model Selection

Choose the right model for the job — don't overspend.

Generation (no input images):

• flux-schnell ($0.003) — Default. Fast, good quality. Use for backgrounds and base scenes.
• flux-dev ($0.03) — Better quality. Use for hero images, portraits, detailed scenes where quality matters.

Editing (with input images):

• seedream ($0.04) — Default. Good for compositing multiple images, placing objects in scenes, adding text. Handles up to 10 inputs.
• banana2 ($0.08) — Better image preservation. Use when you need the input image to stay more faithful, or >10 inputs.
• banana-pro ($0.15) — Best quality, best text rendering. Use for premium work, complex edits, character consistency.

Background removal:

• bria (default) — Best edge quality, clean cutouts
• birefnet — Good general purpose
• pixelcut — Alternative
• rembg — Cheapest

How to Write Good Prompts

This is the difference between mediocre and professional output. Read references/prompt-library.md for a full library of tested prompts you can copy and adapt. Key rules:

For generation: Be specific about lighting ("dramatic side lighting from upper right"), camera ("shot on Canon R5 70-200mm f2.8"), and atmosphere ("dust particles visible in the light beam"). Vague prompts produce generic results.

For text-behind-subject: The key phrase is: *"Add '[TEXT]' in large bold [color] letters BEHIND the [subject] — the [subject's] body overlaps and partially covers the letters."* Without "BEHIND" and the occlusion instruction, the text floats on top.

For edits: Always end with *"Keep everything else exactly the same"* and list what to preserve. Without this, the AI changes things you didn't want changed.

For background replacement: Use realistic, specific locations ("modern upscale mall entrance during daytime, natural warm daylight"). Over-dramatic backgrounds ("city at night with neon reflections") look obviously fake.

Typography

For big titles and hero text: Use the FAL model via edit — it handles large text well and integrates it into the scene naturally. No font size math needed, just say "very large bold" in the prompt.

For precise small text (credits, URLs, badges, coverlines): Use compose or text with Satori. This is where font sizing matters — images display much smaller on phones. Quick rule: on a 1080px Instagram image, nothing under 36px is readable. Run picture-it download-fonts first if fonts aren't installed.

Hierarchy: Max 3 text sizes per image. Brand name should be larger than tagline.

Font pairing: Serif + sans-serif works best. For FAL model text, just describe the style in the prompt. For Satori, 3 fonts are bundled — drop more .ttf files into ~/.picture-it/fonts/. Run picture-it download-fonts if fonts aren't installed. See references/composition-guide.md for pairing suggestions.

Composition Techniques

Read references/composition-guide.md for detailed multi-pass workflows, product photography, magazine covers, and overlay composition.

Common Workflows

Simple: Generate an image

picture-it generate --prompt "dark cosmic background with nebula" --size 1200x630 -o bg.png

Simple: Add text to an image

picture-it text -i bg.png --title "Hello World" --font "Space Grotesk" --color white --font-size 64 -o hero.png

Medium: Blog header with AI background + text

picture-it generate --prompt "abstract dark tech background" --size 1200x630 -o bg.png
picture-it text -i bg.png --title "My Blog Post" --font "DM Serif Display" --font-size 72 -o header.png
picture-it grade -i header.png --name cinematic -o header-graded.png

Medium: Edit a photo background

picture-it edit -i photo.jpg --prompt "replace background with modern hotel entrance, keep subject identical" --model banana-pro -o edited.jpg

Advanced: Text behind subject (YouTube thumbnail style)

# 1. Generate a scene
picture-it generate --prompt "runner on mountain trail at golden hour" --model flux-dev --size 1280x720 -o runner.png

# 2. Use FAL edit to add text BEHIND the subject
picture-it edit -i runner.png --prompt "Add 'RUN FASTER' in large bold black letters BEHIND the runner — the runner's body overlaps the text" --model seedream -o thumbnail.png

Advanced: Product comparison with real photos

# 1. Remove backgrounds from product photos
picture-it remove-bg -i product-a.png --model bria -o a-cutout.png
picture-it remove-bg -i product-b.png --model bria -o b-cutout.png

# 2. Generate a background
picture-it generate --prompt "split gradient, blue left to orange right" --size 1200x630 -o bg.png

# 3. Compose cutouts onto background with text
picture-it compose -i bg.png --overlays overlays.json -o comparison.png

Advanced: Multi-pass cinematic composition

# 1. Generate base scene
picture-it generate --prompt "dark stage with green spotlight" --model flux-dev --size 2048x1080 -o stage.png

# 2. Edit scene to place objects
picture-it edit -i stage.png -i logo.png --prompt "Place Figure 2 as glowing 3D cube in the spotlight" --model seedream -o composed.png

# 3. Post-process
picture-it crop -i composed.png --size 1200x630 --position attention -o cropped.png
picture-it grade -i cropped.png --name cinematic -o graded.png
picture-it vignette -i graded.png --opacity 0.3 -o final.png

Platform Presets

Use --platform <name> with generate or crop:

Output Behavior

• stdout: only the output file path
• stderr: progress logs
• Exit 0 on success, Exit 1 on failure

Read stdout to get the file path. This is how you chain commands.

Gotchas

• Always use --model bria for remove-bg — the default birefnet leaves rectangular artifacts that cause ugly glow/shadow halos when compositing.
• The glow effect in compose mode blurs the entire rectangular buffer, not the shape. Avoid using glow on cutout images — use the background color/lighting to create the glow effect instead.
• The shadow effect has the same rectangular artifact issue. For cutout images on clean backgrounds, skip shadows entirely.
• When editing with FAL, the model may alter product details (logos, text, design elements). For product images where accuracy matters, use remove-bg + compose instead of edit to preserve the original exactly.
• SeedDream takes ~60 seconds per generation. Don't assume it failed if it's slow.
• For edit with banana-pro, don't pass resolution or limit_generations params — it auto-detects.
• Always crop to exact dimensions after FAL generation — FAL models output approximate sizes.
• Use flux-dev ($0.03) not flux-schnell ($0.003) when image quality matters (hero images, portraits). The quality difference is significant.
• Satori does NOT support: display:grid, transforms, animations, box-shadow, filters. Use flexbox only.
• When adding text behind a subject with edit, be very explicit in the prompt: "the text is BEHIND the subject — the subject's body overlaps and partially covers the letters."