lab-image

name	lab-image
description	Generate stunning, on-brand cover images for the AICell Lab site (daily digests, posts, projects) with Cloudflare Workers AI. Use when a post/digest needs a featured image, or to refresh a project cover. Backed by scripts/lab-image.py (dependency-free) with a baked-in signature visual style for consistency.

Lab Image — cover generation

scripts/lab-image.py turns a one-line subject into a polished cover using Cloudflare Workers AI (default model FLUX.1 [schnell] — a strong, fast, recent text-to-image model). Every image carries the AICell Lab signature style so the whole site looks cohesive.

Setup (already configured)

Credentials live in ~/.svamp/.env (never committed): CLOUDFLARE_ACCOUNT_ID, CLOUDFLARE_API_TOKEN (token has Workers AI scope; verified). The CLI auto-loads them.

Commands

scripts/lab-image.py verify                      # check the token
scripts/lab-image.py models                      # list image models
scripts/lab-image.py generate --prompt "<subject>" --out <path.jpg> [--model flux|sdxl|...] [--steps N]
scripts/lab-image.py generate --prompt "…" --out … --raw      # skip the house style
scripts/lab-image.py generate --image <base.jpg> --prompt "…" --out … [--strength 0..1] [--seed N]   # img2img

You pass only the subject (what the cover is about) — the signature style is appended automatically. FLUX returns a 1024×1024 JPEG; the CLI saves it and fixes the extension to match the real format.

The AICell Lab signature style (our visual trail)

Flat, black & white, with orange as the only accent. One recognizable look across the whole site: clean flat minimalist illustration / line-art, strictly monochrome (black on white), with a single bright-orange highlight used sparingly for the active accent — no gradients, no shading, no 3D, no other colors, no text. It's baked into HOUSE_STYLE in scripts/lab-image.py; keep it consistent. For icons, also ask the subject to fill the frame. Examples: the Happy Agent avatar and the newsletter/BioEngine covers.

Use it for posts / digests

Generate the cover into the post folder as featured.jpg; Hugo auto-processes it:

scripts/lab-image.py generate \
  --prompt "<one line capturing the post's topic>" \
  --out content/post/<slug>/featured.jpg

Then make sure the post front matter has an image: block (the templates already do). Keep covers < 1 MB (FLUX JPEGs are ~0.4–0.6 MB — fine).

Image-to-image / identity preservation (mascot variations)

Pass a base image with --image PATH to generate a variation of an existing picture (e.g. our Happy Agent mascot in a new pose) instead of a fresh image from text. The --prompt then guides the change.

# Make the mascot wave hello while keeping its identity (low strength = stays on-model)
scripts/lab-image.py generate \
  --image content/authors/happy-agent/avatar.jpg \
  --prompt "the same cute chibi robot mascot character, now waving hello with one hand" \
  --strength 0.4 --seed 7 --out /tmp/wave.png

Knobs:

• --strength (0..1, default 0.6) — how far to drift from the base. LOWER preserves identity better. In our tests ~0.4 kept the mascot clearly recognizable (antennae, masked face, big eyes, blush, orange flask, silhouette); 0.6 started to lose the face; 0.75 produced a different robot. Start at 0.35–0.45.
• --seed INT — fix it for reproducible runs / A/B comparisons across strengths.

Honest limits (read before promising "same face"):

• There is no true InstantID / IP-Adapter identity-lock model on Workers AI. This img2img mode preserves composition / pose / silhouette, not a guaranteed exact face. Expect a family-resemblance variation, not a pixel-faithful clone. Re-roll seeds / nudge strength down until it looks like us; pick the best of a few.
• Model: only @cf/runwayml/stable-diffusion-v1-5-img2img (sd15-img2img) actually accepts image input on the REST API. FLUX is text-only, and despite the docs SDXL-base rejects image input — so --image with any non-img2img model auto-switches to sd15-img2img (it prints a note). Output is a 512×512 PNG.
• Style drift: SD-1.5 may soften our strictly-flat black-&-white look (faint background texture, muted orange). Mitigate by keeping the house-style prompt (don't use --raw) and using low strength. For a crisp on-brand cover from scratch, prefer plain text-to-image with FLUX.

Matching writing flavor (so words + image feel like one voice)

The digest/post voice is the verbal half of the same identity: warm and first-person-plural, curious and optimistic but grounded, concrete and lightly witty, generous with links, each item a bold headline + 2–4 plain-language sentences, ending with a one-line "why it matters for the lab." Stunning visuals, clear human words — consistently ours.