Gpt Image

GPT Image 2.5 (via DroidProxy Codex)

POST http://localhost:8317/v1/images/generations (or /edits) with Authorization: Bearer dummy-not-used. Send only gpt-image-2.5-flare or gpt-image-2.5-sunburst. Never send an older image model, a dated snapshot, or a chat id (gpt-* without image). Never use an OPENAI_API_KEY. Chat models on :8317 do not generate images.

Default to Flare for fast, high-quality everyday generation. Use Sunburst when the user prioritizes demanding quality, editing precision, or subject preservation over latency. If Sunburst meets the requirements, prefer Flare only after it produces acceptable results on the same prompt, references, dimensions, and quality setting.

Use this for generating or editing visual assets, not for understanding an image the user already attached.

Generate

Requests often take 15–70s. Use --max-time 180. Default quality to auto; select a higher explicit setting only to meet a quality requirement.

RESP=$(curl -sS --max-time 180 http://localhost:8317/v1/images/generations \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer dummy-not-used" \
  -d "$(jq -n --arg p "A red balloon on a wooden table, soft natural light" \
    '{model:"gpt-image-2.5-flare", prompt:$p, size:"1024x1024",
      quality:"auto"}')")

if B64=$(printf '%s' "$RESP" | jq -er '.data[0].b64_json' 2>/dev/null); then
  printf '%s' "$B64" | base64 -d > out.png
else
  printf '%s\n' "$RESP" >&2
  exit 1
fi

Check .data[0].b64_json before decoding. On failure print the raw body — Codex errors are JSON (usage_limit_reached, auth_not_found, moderation_blocked).

Edit

Same auth. JSON with a data URL — no upload step, no file_id. Do not pass the data URL through jq --arg: a typical generated PNG is larger than ARG_MAX (~1MB on macOS) and jq dies with Argument list too long. Write the data URL to a file and use --rawfile, then POST with --data-binary. Decode the same way as Generate.

SRC_FILE=$(mktemp)
REQ=$(mktemp)
printf 'data:image/png;base64,%s' "$(base64 -i ./photo.png)" > "$SRC_FILE"
jq -n --arg p "Make it blue-tinted studio lighting" --rawfile u "$SRC_FILE" \
  '{model:"gpt-image-2.5-sunburst", prompt:$p, images:[{image_url:$u}],
    quality:"auto"}' > "$REQ"

RESP=$(curl -sS --max-time 180 http://localhost:8317/v1/images/edits \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer dummy-not-used" \
  --data-binary @"$REQ")
rm -f "$SRC_FILE" "$REQ"

if B64=$(printf '%s' "$RESP" | jq -er '.data[0].b64_json' 2>/dev/null); then
  printf '%s' "$B64" | base64 -d > out.png
else
  printf '%s\n' "$RESP" >&2
  exit 1
fi

Pick the data-URL mime from the source (image/jpeg, image/png, image/webp). Optional mask.image_url (PNG with alpha) marks the region to replace. Chain edits by feeding each output back as the next images[0].image_url.

Prompting

If the user gives a detailed prompt or asks you to use theirs, use it verbatim. Otherwise define the intended result, then describe the subject, action/pose, setting, style, composition, lighting/mood, visible details, and constraints. For complex work, use labeled sections. Keep requirements easy to read rather than relying on special prompt syntax.

Quote required text exactly, specify its placement and typography, ask for no extra text, and verify spelling and legibility. For edits, say “change only X” and explicitly list what must stay the same. Identify each reference image by number and role (subject, style, clothing, or background). Make one change per edit and restate critical preservation constraints on every turn.

Ground named people, brands, places, and “current/latest” facts with a web search first. For a named real person, edit from a real reference photo — do not generate the likeness from text alone.

Parameters

Response

{"created": 0, "background": "opaque", "data": [{"b64_json": "..."}],
 "output_format": "png", "quality": "low", "size": "1536x1024", "usage": {}}

GPT Image 2.5 always returns base64, so do not send response_format. .data[0] is only b64_json — no url, no mime_type. Pick the file extension from top-level output_format when present, otherwise from the decoded magic bytes (PNG / JFIF / RIFF…WEBP).

Codex OAuth often ignores size and sometimes quality. A 1024x1024 / low request may come back 1536x1024 / low, or an off-enum size like 1312x1199 with quality: medium. Trust the response size / quality / background and the decoded IHDR. Do not retry just because they differ from the request.