Skill: Image generation

Generate images with a local ComfyUI backend (free, private, slow) or a hosted API (fast, paid, better with text inside the image).

Trigger patterns

• Someone asks for an image, visual, or graphic
• A draft or post needs a hero image or social card
• "Generate an image of...", "Make a visual for...", "Create artwork for..."

Choosing a backend

	ComfyUI + FLUX.1 (local)	Hosted API
Speed	minutes, roughly 11 on consumer hardware	seconds
Cost	free after the hardware	per image, billed
Privacy	nothing leaves the machine	prompt and output transit a third party
Text in images	unreliable	good
Best for	background and batch work	quick turnarounds, iteration, text overlays

Default to the local backend for anything not time-sensitive. Reach for the hosted one when the request is urgent, when you are iterating on a design and need the feedback loop tight, or when the image has to contain readable text.

If the request says "quick", "fast", or "now", use the hosted backend.

Both backends cost something. The local one costs wall-clock time and a large chunk of memory - it will evict other models. The hosted one costs money per call. Do not generate speculatively.

Prerequisites

Local (ComfyUI):

• ComfyUI reachable at comfyui_url, running the workflow graph at comfyui_workflow_path
• FLUX.1 weights plus the T5 XXL, CLIP-L and VAE files in ComfyUI's models directory
• comfyui_output_dir set to ComfyUI's own output directory and readable from here

Hosted:

• OPENAI_API_KEY present in the environment, from the chassis secret store. The script reads the process environment only and never opens a credentials file.

Workflow

Step 1 - write the prompt

Both backends respond to descriptive, visual prompts. Cover:

• Subject: what is in the image
• Style: photography, illustration, watercolor, minimal, cinematic
• Lighting: natural light, golden hour, studio, neon
• Composition: close-up, wide angle, overhead, symmetrical
• Mood: warm, energetic, calm, professional

Good:

A bright coworking space with exposed brick walls, developers at laptops, warm afternoon light through tall windows, photography style, shallow depth of field

Bad:

coworking space

Step 2 - generate

# Configured default backend
bash "$IMAGEGEN_SCRIPT" "your prompt here"

# Force local
bash "$IMAGEGEN_SCRIPT" --comfyui "your prompt here"

# Force hosted
bash "$IMAGEGEN_SCRIPT" --openai "your prompt here"

# Explicit output path (works with either backend)
bash "$IMAGEGEN_SCRIPT" --openai "your prompt here" /path/to/my-image.png

The script prints the output path on stdout and progress on stderr, so IMG=$(bash "$IMAGEGEN_SCRIPT" "...") captures the path cleanly.

Default output is <output_dir>/generated-<timestamp>.png.

Step 3 - look at it

Read the generated file and inspect it. If it is wrong, adjust the prompt with more specific descriptors or different style keywords. Iteration is cheap on the hosted backend and expensive on the local one, so be more deliberate about prompt changes when running locally.

Step 4 - use it

Attach it to the reply, or reference the path from whatever draft needed it.

Limitations

• The local backend takes minutes and a large amount of memory. Other local models will be swapped out while it runs.
• The hosted backend costs money per image and needs a working internet connection.
• Neither backend is reliable at rendering a specific string of text, though the hosted one is markedly better.
• Generated images land in the configured output directory. Nothing is cleaned up automatically - that is the install's business.

Prompt tips by use case

Use case	Style keywords
Article hero	cinematic, wide angle, vibrant, professional photography
Social card	minimal, clean background, centered subject, bold colors
Event promo	energetic, crowd, warm lighting, urban setting
Technical post	clean desk setup, code on screen, minimal, soft lighting

Swapping the workflow graph

The shipped graph is FLUX.1 Dev in API format. To use a different one, point comfyui_workflow_path at your own export and set comfyui_prompt_node, comfyui_seed_node and comfyui_output_node to the matching node ids. Those three ids are the entire contract between this script and the graph: prompt in, seed in, filename out.