popcon Dev Log #8 — Archive Page, Start-Frame Review, and BiRefNet VFX Recovery

Thu, 16 Apr 2026 00:00:00 +0900

Overview

Added an archive page to browse past jobs and resume from any pipeline step. Introduced a per-action start-frame review flow so each action’s results can be inspected before moving to refine. Discovered that BiRefNet matting was stripping VFX elements like motion lines and sparks along with backgrounds, and implemented a rescue_vfx_elements() post-processing step to recover them.

Previous post: popcon Dev Log #7 — RunPod GPU Worker, BiRefNet Matting, and Parallel Frame Inference

Archive Page — Browse Past Jobs and Resume

As the pipeline grew longer, revisiting intermediate results or re-running from a specific step became a frequent need. A new archive page solves this.

/archive route displays past jobs as cards
Each card shows current status (which step was last completed)
Selecting a step resumes the pipeline from that point

The layout was updated to include an Archive link in the navigation for easy access.

Per-Action Start-Frame Review Flow

Previously, start-frames for all actions were generated in bulk and reviewed together. This release switches to a per-action flow where each action’s start-frame is reviewed individually, with the option to jump straight into refine if corrections are needed.

flowchart LR
 A["Select Action"] --> B["Generate Start-Frame"]
 B --> C["Tile Review"]
 C -->|Approve| D["Next Action"]
 C -->|Needs Edit| E["Refine Step"]
 E --> C

Key Changes

backend/pipeline/start_frame_gen.py — split start-frame generation to run per action
frontend/components/StartFrameReview.tsx — redesigned tile layout with inline per-emoji video preview
frontend/components/ActionSelector.tsx — integrated action selection UI into the review flow
backend/models.py — extended models for per-action state tracking

Refine resume logic was also improved so that an interrupted refine session can pick up from the last saved state.

BiRefNet VFX Recovery Problem

BiRefNet produces much cleaner background removal than rembg, but it has a blind spot: it classifies VFX elements (motion lines, sparks, speed lines) as background and removes them.

Problem Analysis

VFX element characteristics:

Small non-white blobs
Scattered around the main character
From BiRefNet’s salient object detection perspective, these are “background noise”

rescue_vfx_elements() Implementation

A post-processing function was added to recover VFX elements dropped by BiRefNet matting.

Detect non-white pixel regions in the original image
Identify blobs below a size threshold that were removed by the BiRefNet mask
Re-add blobs likely to be VFX elements back into the mask

Comparative testing of rembg vs. BiRefNet confirmed that BiRefNet + VFX recovery produces the best results.

Start-Frame Tile Redesign

StartFrameReview.tsx was fully redesigned.

Grid tile layout for at-a-glance comparison of each emoji’s start-frame
Inline video preview per tile to immediately check animation results
Per-tile approve/regenerate buttons

Commit Log

Message	Changes
feat(archive): browse past jobs and resume any step	2 files
feat(pipeline): per-action start-frame review + refine resume	12 files
feat(gpu-worker): replace rembg with BiRefNet matting	7 files
feat(review): redesign start-frame tiles and inline per-emoji video	1 file

Insights

BiRefNet’s limitations can be patched with post-processing. Salient object detection models focus on the primary subject, so small VFX elements get lost. A blob-recovery pass is a reusable pattern for any matting pipeline.
Resume capability becomes essential as pipelines grow. Without the archive page, every iteration meant starting from scratch. Persisting intermediate state and allowing re-entry at any step dramatically speeds up development.
Smaller review units make for faster feedback loops. Reviewing all actions at once made it hard to spot issues. Switching to per-action review tightened the iteration cycle.

Pipeline on ICE-ICE-BEAR-BLOG