ComfyUI assets

Use ComfyUI for bitmap generation when the project already has, or explicitly chooses, a local ComfyUI workflow. Prefer the project's existing server, models, custom nodes, and output conventions. Do not install models, custom nodes, or a separate integration merely because this skill was selected.

Keep the server bound to loopback unless the user has requested and secured remote access. Custom nodes and model tooling execute third-party code locally; review their source and provenance before adding them to an environment.

Submit bounded jobs

ComfyUI commonly exposes /prompt, /history/{prompt_id}, and /view. Submit a workflow in API format, retain every returned prompt_id, and bound both each HTTP request and the overall wait:

async function requestJson(url, init = {}) {
  const response = await fetch(url, {
    ...init,
    signal: AbortSignal.timeout(10_000),
  });
  if (!response.ok) throw new Error(`${response.status} ${response.statusText}`);
  return response.json();
}

const baseUrl = "http://127.0.0.1:8188";
const queued = await requestJson(`${baseUrl}/prompt`, {
  method: "POST",
  headers: { "content-type": "application/json" },
  body: JSON.stringify({ prompt: workflow, client_id: clientId }),
});
if (typeof queued.prompt_id !== "string" || !queued.prompt_id) {
  throw new Error("ComfyUI did not return a prompt ID");
}

const deadline = Date.now() + 5 * 60_000;
let result;
while (Date.now() < deadline) {
  const history = await requestJson(
    `${baseUrl}/history/${encodeURIComponent(queued.prompt_id)}`,
  );
  result = history[queued.prompt_id];
  if (result) break;
  await new Promise((resolve) => setTimeout(resolve, 700));
}
if (!result) throw new Error(`Timed out waiting for ${queued.prompt_id}`);
if (result.status?.status_str === "error" || result.status?.completed === false) {
  throw new Error(`ComfyUI job ${queued.prompt_id} failed`);
}

Choose timeout values from the expected local model latency and batch size. A timeout ends the wait, not the server job. Check the owned job's state before retrying so a lost response does not create duplicate generation work. Inspect the returned status/messages when reporting a failure. Encode every /view query parameter and write response bytes only to a chosen output directory; never treat a server-returned filename or subfolder as a trusted local path.

On a shared server, delete only queued prompt IDs submitted by this task. The common /interrupt endpoint affects the active server job globally, so use it only on an isolated server or after confirming that the active job is yours. Keep a record of owned prompt IDs rather than clearing the whole queue or history.

Select models by a local comparison

Inspect what is already installed, then render a small representative comparison when model choice affects quality or cost. Compare prompt adherence, style fit, latency, peak memory, and output consistency on the actual machine. SDXL, FLUX, and their variants have different strengths, but quantization, resolution, sampler, accelerator, and implementation can reverse generic speed or memory claims. Record the chosen model, workflow, seed, sampler, steps, and settings needed to reproduce shipped assets.

Match the asset pipeline

• Runtime-lit materials: request flat, even illumination and reject baked highlights or shadows that fight the runtime light. Baked lighting can be intentional for unlit or stylized pipelines.
• Seamless textures: use a workflow that provides circular/tiling behavior, then inspect a 3×3 composite for seams and distracting repetition.
• Cutouts: generate against a separable background, remove it with an existing segmentation tool, trim transparent margins, and inspect edge halos at final display scale.
• Libraries and variants: hold the art direction and relevant generation settings stable enough for coherence. Vary them deliberately when a local comparison shows that one setup does not serve every asset class.

Generate only enough candidates to make the decision. Use contact sheets for large batches, inspect anomalies and likely rejects closely, and inspect every asset selected for shipping at its intended size and beside its siblings. Check dimensions, alpha, tiling where applicable, value range, compression, and file weight.

Publication and privacy

ComfyUI images can embed the workflow, prompt, model names, node configuration, and source paths in metadata. Before committing, publishing, or sending an asset, inspect the publication copy and remove metadata that is not intentionally part of the deliverable. Keep prompts, workflow JSON, seeds, input images, and model details private unless the user wants them distributed and they have been reviewed for credentials, personal data, private paths, and licensing terms.

Preserve a reproducibility manifest in an approved project location when it is useful, but do not assume it belongs beside public assets. Report which outputs were visually inspected and which were assessed only through a contact sheet or automated checks.