Set up your own image generator

• GitHub — holds your copy of the generator.
• Civitai — where most models come from. Needed if the worker's page asks for a Civitai token.
• RunPod — the computer that makes the pictures. Then Billing → add $10.
  This is a referral link: after your first $10, RunPod gives you a bonus credit (usually $5–10) and Cobalt Pear a small credit.

A «worker» is the model packed for RunPod. Each model has its own worker page on GitHub — its page links here. Made your own worker? Make a pack for it.

1. On the worker's GitHub page, click the green Use this template → Create a new repository.
2. Any name is fine. Choose Private → Create repository.

Keep the worker's own page address — you will paste it into Cobalt Pear in step 7 (not the address of your copy).

No worker yet? Start with a model from the app's Models list (like Z-Image Turbo) — it needs no GitHub at all: skip steps 2–5 and do the box below instead.

1. Open Cobalt Pear → Models → Add next to the model. The RunPod steps show in the same window.
2. Do those steps on RunPod — for Z-Image Turbo: Serverless → New Endpoint → Deploy image or video API → Text To Image → Tongyi-MAI/Z-Image-Turbo → Deploy with vLLM Omni → Next → Create Endpoint.
3. In the endpoint's settings set Max workers 1, Execution timeout 600 → Save.
4. Paste its endpoint ID next to the model in Cobalt Pear. Then do step 6 and put the key in Settings.

No build to wait for. The first picture takes about a minute while RunPod starts it; then about 8 seconds each.

Only if the worker's page asks for it. It lets RunPod download the model with your account.

1. Civitai → your picture (top right) → Account settings → Security & Apps → API Keys → Add API key → Permission preset Read Only → copy it and keep it somewhere — Civitai shows it only once. One token works for every model: next time, reuse it.
2. In your copy on GitHub, click Dockerfile → the pencil ✏️ (top right of the file).
3. Replace PASTE_YOUR_CIVITAI_TOKEN_HERE with your token → Commit changes… → Commit changes again.

The «endpoint» is your generator on RunPod.

1. RunPod → Serverless → New Endpoint → Deploy from a GitHub repository.
2. Pick your copy. The first time, click Edit connection and let RunPod see it.
3. Fill in the page from top to bottom. Some settings are folded away — the arrows show what to click open:

4. Leave the rest as it is → Deploy at the bottom.

RunPod may warn «runpod.serverless.start() handler not found in your repo». You can ignore it — the worker starts its handler from the image it builds on.

RunPod now builds your generator — about 30 minutes. The endpoint's page shows its progress. You can do step 6 meanwhile.

If it fails, open the build log, copy the last lines and ask in Discussions.

1. RunPod → Settings → API Keys → Create API Key.
2. Choose All.
3. Copy the key now — RunPod shows it only once.

One key does everything — making pictures and making endpoints — for every endpoint you add later. Cobalt Pear keeps it in your browser, so use a browser only you use. RunPod spends only the credit you added.

1. Open Cobalt Pear → Models → paste the worker's page address (from step 2) under «Your own worker» → Add.
2. Paste your endpoint ID next to the model — on RunPod it is the short code under the endpoint's name (like ab12cd34ef56gh), then go to Chat.
3. Type what you want to see, press Enter. The first time it opens Settings — paste your RunPod API key from step 6 there, then go back to Chat.

The very first picture can take 10–15 minutes while RunPod starts the generator for the first time. After that it is seconds to a minute.

A LoRA: Models → + LoRA under the model. The first time, pick how that model gets its LoRAs:

• Built into the worker — for LoRAs you keep using. Paste the links; the app shows one line to put in your copy's Dockerfile (with a link that opens it) and RunPod builds the worker again. No waiting at each picture, and you pick which ones to use for each picture from the LoRA button next to Send. There is no set number of LoRAs — the limit is time: RunPod stops a build that runs over 30 minutes, and each LoRA makes the build longer and the worker bigger (the app shows how many GB they add), so keep it to the ones you use. Changing the list means building again; until it's built, pictures come out without the new LoRAs and say so. Nothing to press when it's done — the next picture tells the app.
• By link — for trying one out. Paste the link and send: the worker downloads it with the next picture (a few seconds more, and again after the worker has been idle). Nothing to build. Up to 4 per model — for more, use «Built in» (or, on RunPod: your endpoint → Manage → Edit Endpoint → Environment Variables → add REMOTE_LORA_MAX = 8 → Save Endpoint).

You can switch any time with the «LoRAs: Built in | By link» switch under the model in Models. Pick a LoRA made for that model (its base model shows when you point at it); one made for a different model does nothing, and the picture then says so with ⚠. LoRAs work on models from a worker repo or «Any Civitai SDXL model» — not on the ones RunPod runs itself (the Hugging Face list). An SDXL model in that list has a Custom button: it runs the same model on your own copy of «Any Civitai SDXL model» instead, so LoRAs work — no Civitai token needed. Only SDXL models get it — that worker runs SDXL, so FLUX, SD 3 and the like (most of the Gated ones) have no Custom. A gated SDXL model (you must agree to its terms on Hugging Face first) adds one step: agree on its page, make a Read token, and paste it as HF_TOKEN= — the setup steps link both. If Civitai won't hand a LoRA over, a Civitai token box shows in Models — paste a token there (Civitai → your picture → Account settings → Security & Apps → API Keys → Add API key → Read Only).