127 sprites had been generated and about eight had ever been looked at. The pipeline will happily render the wrong object, confidently and well-lit, and no assertion can tell — barrier and flightCase are both "a grey box with edges" as far as any test goes. tools/gen/contact_sheet.py puts every shipped sprite on one page at integer upscale on the venue's own floor colour. It immediately showed about twenty props reading as pale blobs. First instinct was that the vertex baker had lost the albedo. That was wrong: sat=0 on ashUrn, marbleSink and pillar is honest, because stainless and marble and concrete really are grey. The fault was BRIGHTNESS. Props that read sit at median luminance 23-123 (patioHeater 23, arcade 45, poolTable 53, cdj 68); the blobs were 150-190. The venue is painted at ~35 under a 50% black sheet and props are lit by the room's own LIGHTS, so a prop arriving at 180 doesn't read as a bright object — it reads as self-illuminated. Every prop that worked before worked by accident of subject matter; the first pale subjects exposed the gap. props_import --dim scales RGB until median opaque luminance is <=110. Only ever darkens, so anything already in band passes through untouched. Fixed 48 sprites with no regeneration at all — the 512px Blender renders were still in art_incoming, which is exactly why that directory is kept. Also: poster7 shipped with 29% of its pixels and poster9 with 42%, because the near-black-to-alpha pass was eating the dark half of a dark poster. Posters are rectangular plates on a wall, so 4-9 now keep their background and fill the frame. And cf_flux --probe was reporting EXHAUSTED on an account with a full 10,000 neurons: the probe prompt was "a grey square", Cloudflare's safety filter refused it as NSFW, and every non-429 fell through to SystemExit. A health check that can't tell "refused" from "empty" will eventually call a healthy system dead. Gate: lint clean, build clean, 821 tests passing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
137 lines
5.3 KiB
Python
137 lines
5.3 KiB
Python
#!/usr/bin/env python3
|
|
"""cf_flux.py — FLUX-1-schnell images from Cloudflare Workers AI.
|
|
|
|
python3 tools/gen/cf_flux.py <out.png> "<prompt>" [--steps 6] [--seed 42]
|
|
python3 tools/gen/cf_flux.py --probe # is there budget left today?
|
|
|
|
Workers AI gives 10,000 Neurons free every day (resets 00:00 UTC), which is the
|
|
cheapest image generation on the fleet — so it goes FIRST, ahead of the farm.
|
|
When it runs dry, callers fall back to MODELBEAST `flux_local`, which is slower
|
|
but genuinely unlimited.
|
|
|
|
The whole reason this exists rather than the one-liner in toastsim: urllib
|
|
raises HTTPError on 4xx and throws the response BODY away, so a 429 is
|
|
indistinguishable between "you have burned today's neurons, go to the farm" and
|
|
"you sent three at once, wait a second". Those want opposite responses, and
|
|
guessing wrong either wastes the free tier or wedges the batch. Here the body is
|
|
read and classified.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import base64
|
|
import json
|
|
import sys
|
|
import time
|
|
import urllib.error
|
|
import urllib.request
|
|
from pathlib import Path
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parent))
|
|
import creds # noqa: E402
|
|
|
|
MODEL = "@cf/black-forest-labs/flux-1-schnell"
|
|
# schnell is a distilled 4-step model; past ~8 steps you pay neurons for nothing.
|
|
DEFAULT_STEPS = 6
|
|
|
|
|
|
class Exhausted(RuntimeError):
|
|
"""Today's free neurons are gone. Fall back to the farm; retrying won't help."""
|
|
|
|
|
|
class RateLimited(RuntimeError):
|
|
"""Too fast, not too much. Backing off DOES help."""
|
|
|
|
|
|
def _post(prompt: str, steps: int, seed: int) -> bytes:
|
|
account = creds.get("CLOUDFLARE_ACCOUNT_ID")
|
|
token = creds.get("CLOUDFLARE_API_TOKEN")
|
|
req = urllib.request.Request(
|
|
f"https://api.cloudflare.com/client/v4/accounts/{account}/ai/run/{MODEL}",
|
|
data=json.dumps({"prompt": prompt, "steps": steps, "seed": seed}).encode(),
|
|
headers={"Authorization": f"Bearer {token}", "Content-Type": "application/json"},
|
|
)
|
|
try:
|
|
raw = urllib.request.urlopen(req, timeout=180).read()
|
|
except urllib.error.HTTPError as exc:
|
|
body = exc.read().decode(errors="ignore")[:600]
|
|
low = body.lower()
|
|
if exc.code == 429:
|
|
# Cloudflare says "capacity temporarily exceeded" for burst limits and
|
|
# names the neuron/daily allocation when the free tier is spent.
|
|
if "neuron" in low or "daily" in low or "limit" in low and "rate" not in low:
|
|
raise Exhausted(body) from None
|
|
raise RateLimited(body) from None
|
|
if exc.code in (401, 403):
|
|
raise SystemExit(f"Cloudflare rejected the token (HTTP {exc.code}): {body}")
|
|
raise SystemExit(f"Cloudflare HTTP {exc.code}: {body}") from None
|
|
|
|
payload = json.loads(raw)
|
|
if not payload.get("success"):
|
|
errors = json.dumps(payload.get("errors"))
|
|
if "neuron" in errors.lower() or "limit" in errors.lower():
|
|
raise Exhausted(errors)
|
|
raise SystemExit("Cloudflare error: " + errors)
|
|
return base64.b64decode(payload["result"]["image"])
|
|
|
|
|
|
def generate(out: Path, prompt: str, steps: int = DEFAULT_STEPS, seed: int = 0,
|
|
*, attempts: int = 4) -> Path:
|
|
"""One image to `out`. Retries burst limits, gives up fast on exhaustion."""
|
|
delay = 2.0
|
|
for attempt in range(1, attempts + 1):
|
|
try:
|
|
out.parent.mkdir(parents=True, exist_ok=True)
|
|
out.write_bytes(_post(prompt, steps, seed))
|
|
return out
|
|
except RateLimited:
|
|
if attempt == attempts:
|
|
raise
|
|
time.sleep(delay)
|
|
delay *= 2
|
|
raise RateLimited("exhausted retries")
|
|
|
|
|
|
def probe() -> bool:
|
|
"""Cheapest possible call — is there budget left right now?
|
|
|
|
The question is ONLY "has the daily allocation run out". Anything else the
|
|
API says means the account is alive and spending is possible, so it counts as
|
|
budget available — including a rejected prompt. That distinction is not
|
|
hypothetical: Cloudflare's safety filter rejected the literal prompt "a grey
|
|
square" as NSFW (HTTP 400, code 3030), and because the old probe let every
|
|
non-429 fall through to SystemExit, a perfectly healthy account with a full
|
|
10,000 neurons reported as unusable.
|
|
"""
|
|
try:
|
|
_post("a plain wooden chair, product photo", 4, 1)
|
|
return True
|
|
except Exhausted:
|
|
return False
|
|
except RateLimited:
|
|
# Burst-limited means the account is alive and has budget.
|
|
return True
|
|
except SystemExit as exc:
|
|
# A 4xx that is not the daily cap: the account works, this prompt did not.
|
|
print(f"cloudflare reachable, prompt refused: {exc}", file=sys.stderr)
|
|
return True
|
|
|
|
|
|
def main() -> None:
|
|
if "--probe" in sys.argv:
|
|
ok = probe()
|
|
print("cloudflare flux: BUDGET AVAILABLE" if ok else "cloudflare flux: EXHAUSTED for today")
|
|
sys.exit(0 if ok else 3)
|
|
if len(sys.argv) < 3:
|
|
sys.exit(__doc__)
|
|
out, prompt = Path(sys.argv[1]), sys.argv[2]
|
|
steps = int(sys.argv[sys.argv.index("--steps") + 1]) if "--steps" in sys.argv else DEFAULT_STEPS
|
|
seed = int(sys.argv[sys.argv.index("--seed") + 1]) if "--seed" in sys.argv else 0
|
|
try:
|
|
print(generate(out, prompt, steps, seed))
|
|
except Exhausted as exc:
|
|
sys.exit(f"cloudflare free tier exhausted for today — use MODELBEAST flux_local.\n{exc}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|