GPT Image 2.6 is useful for more than producing a single attractive image. A harder developer problem is generating a set of animated assets while keeping one character recognizable across poses, expressions, and motion.

This tutorial turns that problem into a two-stage pipeline:

  1. generate and approve a static character master;
  2. edit that approved master into native GIF variants, then validate both the file and the visual invariants.

A sticker pack is the running example, but the same design applies to animated avatars, onboarding mascots, and reaction icons. The code below is a minimal unexecuted example—use placeholder credentials and test outputs before integrating it into production.

Why generation and animation should be separate

Trying to define identity, style, pose, motion, timing, and format in one prompt creates too many degrees of freedom. Instead, make the first output a contract.

A useful master sheet contains one character on a plain background plus a short written specification:

{ "character": "round capybara mascot", "must_keep": [ "small rounded ears", "dark oval nose", "blue neck scarf", "warm brown fur", "thick dark outline" ], "may_change": ["pose", "expression", "small props"] } 

Review this static asset before creating animation. That checkpoint prevents a weak design from multiplying into a whole batch.

For exploratory character sheets, start with GPT Image 2.6 Flare. When the task becomes a tightly constrained edit around a reference, compare it with GPT Image 2.6 Sunburst. Treat that split as a workflow heuristic, not a universal quality guarantee: evaluate both against your own acceptance set.

If you prefer a visual workflow before writing code, the same generate-select-edit loop can be prototyped in Ace Data Cloud Studio.

Stage 1: create the master

The public Images Generations API documentation exposes an OpenAI-compatible endpoint. This example intentionally requests a static PNG rather than a GIF:

const API_BASE = "https://api.acedata.cloud"; const token = process.env.ACE_API_TOKEN; // placeholder; do not hard-code const masterResponse = await fetch(`${API_BASE}/openai/images/generations`, { method: "POST", headers: { Authorization: `Bearer ${token}`, "Content-Type": "application/json" }, body: JSON.stringify({ model: "gpt-image-2.6-flare", prompt: [ "Create a clean character master for a round capybara mascot.", "Warm brown fur, small rounded ears, dark oval nose,", "blue neck scarf, thick dark outline, plain background.", "Front three-quarter view; no text; no extra characters." ].join(" "), size: "1024x1024", quality: "high", output_format: "png", n: 1 }) }); if (!masterResponse.ok) { throw new Error(`generation failed: ${masterResponse.status}`); } const master = await masterResponse.json(); const masterUrl = master.data?.[0]?.url; if (!masterUrl) throw new Error("response contained no image URL"); 

Store the result in your own object storage with a content hash, prompt version, model ID, and approval status. A provider URL is a delivery mechanism, not a durable asset-management strategy.

Stage 2: edit the approved reference into a GIF

The Images Edits API documentation accepts image URLs in JSON mode. GPT Image 2.6 Flare and Sunburst support GIF output, including frame-count and frame-duration controls.

const editResponse = await fetch(`${API_BASE}/openai/images/edits`, { method: "POST", headers: { Authorization: `Bearer ${token}`, "Content-Type": "application/json" }, body: JSON.stringify({ model: "gpt-image-2.6-sunburst", image: masterUrl, prompt: [ "Animate only the approved capybara character waving once.", "Preserve the ears, nose, blue scarf, fur color, outline style,", "body proportions, and plain background.", "Create a clean loop; add no text and no new character." ].join(" "), output_format: "gif", size: "512x512", frame_count: 9, frame_duration_ms: 200, n: 1 }) }); if (!editResponse.ok) { const diagnostic = await editResponse.text(); throw new Error(`edit failed: ${editResponse.status} ${diagnostic}`); } 

The documented GIF controls allow 2–16 frames and frame durations from 20 to 655,350 ms. Encoding rounds duration to the nearest 10 ms. Most importantly, static 4K support does not imply 4K GIF support; 512x512, 9 frames, and 200 ms are documented preview-oriented defaults.

Validate before accepting the result

A 200 response only proves that the request completed. Add two acceptance layers.

First, validate the artifact mechanically:

  • response contains exactly one usable URL;
  • downloaded bytes have the expected GIF signature and MIME type;
  • dimensions, frame count, duration, and file size fit your application;
  • the asset is copied to durable storage and recorded idempotently.

Second, validate preservation. A reviewer—or a vision-based test with human fallback—should score each must_keep feature and reject unexpected text, extra characters, broken loops, or background drift. Save the rejection reason; the next edit prompt should target the failed invariant rather than regenerate everything.

For queues, use one idempotency key per (master_hash, action, prompt_version). Retry transport failures and rate limits with backoff, but do not blindly retry validation failures. Authentication failures require operator intervention; moderation refusals require changing the request, not hammering the endpoint.

This separation gives you an auditable workflow: identity is approved once, motion is varied deliberately, and each output has explicit technical and visual acceptance criteria.

Sources and disclosure

The workflow was adapted from Ace Data Cloud's public article, Use Ace Data Cloud Studio to create a set of your own WeChat animated stickers, and the public generation/edit API documentation linked above. The source article's capybara process informed the example, but this post does not claim that the current author personally ran it or that any sticker pack was submitted, approved, or listed by WeChat.

This is promotional content for Ace Data Cloud, produced with AI assistance and editorial adaptation from Ace Data Cloud materials. The API snippets are illustrative and were not executed for this article.