Spark Comfy

Spark Atlas — Fleet Image Generation

Spark Atlas is the Weblyfe fleet's on-prem DGX Spark inference box (Tailscale-only endpoint set by SPARK_BASE). It runs ComfyUI behind a FastAPI wrapper called spark-api on port 8190. Any Appie on the tailnet can render images by hitting it.

This skill is the client side: a thin Python helper and copy-paste curl examples.

Prerequisites

1. Be on the same Tailnet as Spark Atlas. Test with:

   curl -s $SPARK_BASE/health | jq .
   

   Should return {"status": "healthy", "comfyui": "connected", ...}.

2. Have an API key. Set it in your ~/.hermes/.env (or your shell env):

   SPARK_API_KEY=spark_<your-handle>_<hex>
   

   Seyed provisions keys via Spark's ~/.hermes/.env SPARK_API_KEYS list. Ask if you don't have one.

3. The Python helper imports the standard library only (urllib, json, time, os). No extra installs needed.

Quick reference

max_concurrent = 2. Submit a third while two are running → 429 Max concurrent jobs reached.

Quick start (curl)

# Health
curl -s $SPARK_BASE/health | jq .

# Generate with SDXL
JOB=$(curl -s -X POST $SPARK_BASE/generate \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $SPARK_API_KEY" \
  -d '{
    "prompt": "luxury portrait of a confident woman, premium editorial",
    "model": "sdxl",
    "style": "cinematic"
  }' | jq -r .job_id)

# Poll until done
while true; do
  STATUS=$(curl -s "$SPARK_BASE/status/$JOB" | jq -r .status)
  echo "$STATUS"
  [[ "$STATUS" == "completed" || "$STATUS" == "failed" ]] && break
  sleep 3
done

# Get the image URL
curl -s "$SPARK_BASE/status/$JOB" | jq -r '.image_urls[0]'
# → /output/job_xxx_SparkAtlas_..._00001.png

# Download it
curl -s -o out.png "$SPARK_BASE/output/job_xxx_..._00001.png"

Quick start (Python)

from spark_client import spark_generate

urls = spark_generate(
    prompt="luxury Dubai penthouse balcony at golden hour, panoramic city view",
    model="flux_schnell",  # or "sdxl"
    style="luxury",
)
print(urls[0])

The spark_client.py helper lives next to this SKILL.md. Copy it into your skill's tools dir or import directly.

Available models

Call /models to see the live list — new models added on Spark show up there automatically.

Style presets

The style field appends a style tag block to your prompt. Available:

cinematic · natural · studio · fashion · luxury · vintage · minimalist · (and more — call /styles).

Set style: "" to skip preset and pass tags inline.

Generate request schema

{
  "prompt": "subject + scene description",
  "negative_prompt": "low quality, blurry, deformed",
  "model": "sdxl",
  "style": "cinematic",
  "width": 1024,
  "height": 1024,
  "steps": 24,
  "cfg": 5.0,
  "sampler": "euler",
  "scheduler": "normal",
  "seed": 42,
  "return_url": true
}

Only prompt is required.

Operating rules

1. Read /system first when something feels off. Tells you active_jobs, available_models, total_generated.
2. Cap to 2 concurrent jobs. Spark's max_concurrent = 2. Queue locally above that.
3. One render → review → iterate. Don't blast 20 variants in parallel — wastes GPU minutes.
4. Right model for the job. flux_schnell for fast iteration; sdxl for the final delivery.
5. Save winners. Persist prompt + seed + model together. Reproducible.

Error reference

Files in this skill

• SKILL.md — this file
• spark_client.py — Python helper (stdlib only)
• examples/ — sample prompts + workflows (TBD)

See also

• docs/CLAUDE-CODE-PATH.md in the appie-kit for general Claude Code orchestrator setup.
• $SPARK_DOCS_DIR/spark-api-instructions.md on the Spark host.
• Operator discipline doc: $SPARK_DOCS_DIR/operator-discipline.md.