Image Gen
Four free backends in a fallback chain, best first. generate.sh walks them in
order and stops at the first that returns a valid image, so a quota exhaustion
or rate limit on one tier never blocks a request.
# defaults to ./img.jpg, walks all four tiers
./generate.sh "a red barn in a snowy field"
# explicit output, named extension is fine - format follows the backend
./generate.sh "minimalist app icon, indigo and cream" -o /tmp/icon.png
# skip to a specific tier (debugging, or forcing the unlimited one)
./generate.sh "a lighthouse at dusk" --tier 3
./generate.sh --list
Use the script. Don't hand-roll the curl calls — every tier has a non-obvious response shape and each has cost me a debugging cycle.
If you have no shell access to this script
You are probably a read-only subagent (explorer, architect, code-reviewer,
verifier, security-reviewer, database-reviewer, docs-lookup). Those agents
deny shell * with an allowlist that excludes this script, so you cannot
generate the image yourself, no matter that the skill loaded fine.
Do not silently skip the task and do not report failure. Emit the exact command and let the parent run it, then read the resulting file:
~/.config/opencode/skills/image-gen/generate.sh "<prompt>" -o <output-path>
If shell is blocked, say which tier you would have started at (default 1) and
why you could not continue. The parent owns asset generation by design: it is the
only context with both shell and the ability to report a verified file back.
Tier order
| Tier | Model | Access | Limit | Resolution |
|---|---|---|---|---|
| 1 | agnes/agnes-image-2.0-flash | OmniRoute Gateway | unmetered | 1024×1024 |
| 2 | stable-diffusion-xl-base-v10 | OVH Kepler AI (anonymous) | 2/min → ~2,880/day | 1024×1024 |
| 3 | @cf/black-forest-labs/flux-1-schnell | Cloudflare Workers AI | 10,000 neurons/day → ~230/day | 1024×1024 |
| 4 | AI Horde (Realistic Vision default) | AI Horde | no daily cap | 512×512 |
| 5 | Pollinations sana | anonymous | ~1 per 15-45s | 768×768 |
Tier order reflects measured output speed and fidelity. Tier 1 (OmniRoute Agnes Image 2.0 Flash) leads because it renders 1024×1024 PNGs in ~5s with prompt fidelity and no watermark. Tier 2 (OVH SDXL) and Tier 3 (Cloudflare FLUX) provide immediate fallbacks.
Measured limits — read before promising anything
Tier 1, OmniRoute. High quality Agnes Image 2.0 Flash via local gateway (http://127.0.0.1:20128/v1/images/generations). Renders 1024×1024 PNGs in ~5-8s. Direct URL retrieval.
Tier 2, OVH. Anonymous, no key, no signup, no card. Headers say it exactly:
x-ratelimit-limit-minute: 2
retry-after: 5
2/min → ~2,880/day. Takes 10-28s. Accepts negative_prompt,
num_inference_steps, guidance_scale. Returns base64 PNG. Undocumented
and carries no SLA — it could vanish without notice.
Tier 2, Cloudflare. 10,000 neurons/day, resets 00:00 UTC. Neuron formula is
tiles × 4.8 + steps × 9.6 × tiles. ~230 images/day at 512×512, ~57 at
1024×1024. Returns base64 JPEG despite the "flux" model name. The API does
not echo the seed back — record it yourself if you need reproducibility.
Tier 3, AI Horde. Volunteer network, no daily cap whatsoever, so this is
the true unlimited tier. Anonymous users are capped at ~580×580, so the script
requests 512×512. Quality is SD-1.5 era: expect geometry artifacts and occasional
false-positive NSFW blocks. Override with HORDE_MODEL=... ./generate.sh "...".
Tier 4, Pollinations. Weakest tier. See the corrections below.
Corrections to earlier versions of this skill
Three things in the old doc were wrong. They are fixed in the script:
- Cloudflare's FLUX endpoint rejects
width/heightwithAiError: Bad input: Additional or unevaluated properties. The old doc claimed they were optional. Omit them. - Pollinations is not FLUX.
https://image.pollinations.ai/modelsreturns exactly["sana"]. Passingmodel=fluxsilently returns a Sana image. - Pollinations watermarks even with
nologo=true. Confirmed by inspecting output. Treat tier 4 as last-resort only.
Verified dead ends
- Google Gemini image models — 429
free_tier_requests limit: 0on a new key. - fal.ai — account billing-locked (
TOP_UP). - Together AI — free tier withdrawn; accounts read-only until you deposit.
- HuggingFace Inference — only ~$0.10/month in credits.
- Local generation — Intel integrated graphics, 7.6GB RAM, no CUDA. Stable Diffusion needs a dedicated NVIDIA GPU.
Quality expectations
These are the bottom of the leaderboard, not the middle. Current frontier (GPT-Image-2.5, Nano Banana 2.1, Seedream 5.0 Pro) scores ~1,425 on Arena text-to-image; FLUX.1 schnell scores ~741. Roughly a 97% head-to-head gap.
Fine for: UI mockups, visual direction, placeholder and fixture images, icon and logo concepts, README hero art, decorative illustration.
Not for: pixel-exact reproduction, brand consistency across a set, legible in-image text (no free tier here renders text correctly), or CI assets needing reproducibility. For UI work, build the real thing in HTML/CSS — generate images to explore direction, then implement in code.
An undocumented alternative worth testing if you want a quality jump: Pollinations
claims flux is free forever on a registered key (no per-IP limit). Unverified
here, and their recent paywall tightening makes it worth probing before trusting.