Previous: PopCon Dev Log #1

Overview

Today’s session split between polishing the “outside” and hardening the “inside” of PopCon. The morning was branding — turning a Gemini-generated image into logo and favicon assets, then writing a GitHub-ready README. The afternoon turned into a Docker debugging session that led to pipeline quality improvements, finishing with retry logic and per-emoji error handling.

1. Logo & Favicon — From Gemini Image to Brand Assets

The first task was converting a 2880×1440 Gemini-generated image into PopCon brand assets. The image was center-cropped to 1:1 then resized into multiple sizes.

Why the Favicon Wasn’t Showing in Docker

Two issues stacked on each other.

1. Next.js App Router priority: app/favicon.ico takes precedence over public/favicon.ico. A default favicon was already sitting in app/ and needed to be replaced there.
2. Docker image baking: The Dockerfile uses COPY . . at build time. Changing files on disk has no effect until the container is rebuilt.

docker compose build frontend && docker compose up -d frontend

Confirmed with curl -I localhost:3000/favicon.ico returning HTTP 200.

2. Full Product README

Next came writing a README that reads like a product landing page rather than a bare technical doc.

The first version put both English and Korean in a single file. Feedback: “the two languages aren’t distinguishable.” Split them into separate files with language toggle links at the top of each.

• README.md — English, with English | [한국어](README.ko.md) at the top
• README.ko.md — Korean, with [English](README.md) | 한국어 at the top

README vs. Reality

After the first commit, reviewing the actual code revealed several discrepancies.

Both README files updated and pushed.

3. Docker Debugging — Fighting the API Key

The afternoon session opened with Docker logs showing nothing working. The root cause was a trailing venv appended to the API key in .env.

POPCON_GOOGLE_API_KEY=AIzaSy...-mAcuv venv

Likely copied from a terminal where the venv activation command ran next. Removed the trailing text, restarted — but the key itself turned out to be expired too. Generated a fresh one from Google AI Studio.

Running the pipeline for real surfaced a series of quality problems.

Issues Found and Fixed

flowchart TD
    A["Pipeline run"] --> B["Review output"]
    B --> C["Problems found"]
    C --> D["Cloud artifacts <br/> in background"]
    C --> E["Duplicate characters <br/> sprite sheets"]
    C --> F["Checkerboard background <br/> NB2 fake transparency"]
    C --> G["VEO edge lines <br/> left/right borders"]
    D --> H["Remove rembg entirely <br/> brightness-based crop instead"]
    E --> I["Prompt: single character only <br/> no sticker sheets"]
    F --> J["Prompt: #FFFFFF background <br/> NOT checkerboard"]
    G --> K["ffmpeg 2% edge crop"]

The biggest decision was removing rembg entirely. Every attempt at background removal made things worse — isnet-general-use left cloud artifacts, u2net wasn’t better. Instead: prompt VEO to generate white backgrounds, then use brightness-based cropping to extract content.

# processor.py — brightness-based content detection
brightness = arr.astype(float).mean(axis=2)
content_mask = (brightness > 10) & (brightness < 245)

Removing rembg[cpu]>=2.0.0 from pyproject.toml and replacing it with numpy>=1.26.0 also slimmed the Docker image.

LINE File Naming Fix

Checking the LINE Creators Market guidelines: files must be named 001.png through 040.png. The code was saving them by action name.

# packager.py
for i, emoji_path in enumerate(emoji_paths):
    line_name = f"{i + 1:03d}.png"
    zf.write(emoji_path, line_name)

4. Retry Logic and API Throttling

The final commit focused on resilience. Generating a full 24-emoji set means many API calls to VEO and Gemini — and they occasionally return 503 or 429. Previously, one failure killed the whole job.

Per-Emoji Error Handling

The fix wraps each emoji’s pipeline stages in a try/except, marks failures with "error" status, and continues with the rest.

# worker.py — per-emoji error handling
failed_indices = set()
for i, action in enumerate(actions):
    try:
        # ... pose generation, animation, post-processing
    except Exception as e:
        logger.error(f"Emoji {i} ({action.name}) failed: {e}")
        failed_indices.add(i)
        status.results[i].status = "error"
        save_job(status)

A new "done_with_errors" job status means the ZIP is still available even if some emojis failed.

API Retry with Exponential Backoff

Both the Gemini Image and VEO calls now retry up to three times on transient errors.

# pose_generator.py — retry logic
async def _generate_image(self, prompt, reference_image_path=None, max_retries=3):
    for attempt in range(max_retries):
        try:
            response = await asyncio.to_thread(
                self.client.models.generate_content, ...
            )
            return ...
        except (ServerError, ClientError) as e:
            if attempt == max_retries - 1:
                raise
            wait = 2 ** attempt  # 1s, 2s, 4s
            logger.warning(f"Attempt {attempt+1} failed, retrying in {wait}s: {e}")
            await asyncio.sleep(wait)

Type System Sync

Status types updated across backend and frontend.

Summary

The four commits today in order:

The Docker debugging session was unexpectedly productive — it forced a real pipeline run that surfaced quality issues, and the decision to remove rembg entirely turned out to be the right call. Less code, smaller Docker image, cleaner output.

Next up: final quality validation before submitting to LINE Creators Market.