Reviving Praderas (Day 17) — cover images: ComfyUI plan + SDXL template

Published on 05/10/26 19:45:00
Web development Productivity Artificial intelligence
Reviving Praderas (Day 17) — cover images: ComfyUI plan + SDXL template

Reviving Praderas (Day 17) — cover images without roulette (plan + ComfyUI)

After the per-language sitemap work (Day 16), we picked up Priority 2 in proposed-improvements.md: replace random placeholder imagery with something predictable or controlled, without forcing a different editorial flow than we already use (Markdown + front matter).

What we decided in this session

  1. Two complementary tracks

    • Deterministic in-repo: keys like Cover: mapped to static assets (no GPU; great when agents only emit YAML).
    • ComfyUI (local): generate a raster from Title / Description + a fixed house-style block, then commit the image and reference it (e.g. Image:) when publishing.

  2. Real technical validation
    Against http://127.0.0.1:8188 we exercised the usual flow: POST /prompt → GET /history/{prompt_id} → GET /view?.... A minimal SD 1.5 @ 512² graph proved wiring; an SDXL “ubersimple” graph (SDXL/sd_xl_base_1.0.safetensors, 1024×768, euler + sgm_uniform, long negative, VAEDecode with the checkpoint VAE ["1",2]) produced a clearly better image — consistent with the intuition that quality was about the full stack, not “VAE alone.”

Example output (same session, local smoke test)

Committed file: assets/images/day17-comfyui-sdxl-example.webp (1024×768). Generic positive prompt (editorial desk / terminal mood, no readable text); meant only to show the quality bar of the SDXL graph above, not as the final cover for a specific post. From Day 18 onward the same path is referenced in front matter as Image: and rendered as the hero above the body (no duplicate ![](...) in Markdown).

What stayed out on Day 17 (and shipped next)

The original publication of this note had no Pico wiring for hero or social metadata; Day 18 adds Image:, og:image, large Twitter card, and responsive layout — see the next series entry (adjacent Translation_Key in the tracker).

Where agents should look

  • .agents/comfyui-cover-images.md — prerequisites, API flow, node table, security notes (localhost vs tunnel), integration checklist (assets, front matter, Twig, script, CI, lint).
  • scripts/comfyui/sdxl_ubersimple.api.json — template JSON: replace node 3 inputs.text, tune seed / SaveImage prefix.

We also linked from repo-context.md, post-template.md (Image: field), and proposed-improvements.md so discovery does not depend on chat history.

Next steps (after Day 18)

  1. Pick Image: + path lint — done on Day 18.
  2. Hero + og:image + responsive CSS — done on Day 18.
  3. Small script (Python) that reads a .md, fills the ComfyUI JSON, saves PNG under assets/ (or Git LFS).
  4. Decide whether generation stays local-only or runs in CI against a reachable instance with secrets.

Wall clock (order of magnitude)

Docs + template + this log + PR: ~25–40 minutes in one pass; the earlier ComfyUI smoke tests were a separate session.