Skip to content

Codex/ollama remote qwen25vl#1

Merged
renatobo merged 7 commits into
mainfrom
codex/ollama-remote-qwen25vl
Mar 19, 2026
Merged

Codex/ollama remote qwen25vl#1
renatobo merged 7 commits into
mainfrom
codex/ollama-remote-qwen25vl

Conversation

@renatobo
Copy link
Copy Markdown
Owner

Summary

Describe what changed and why.

Changes

Validation

  • make lint
  • make test

Checklist

  • Tests added/updated for behavior changes
  • Docs updated (README/docs) if needed
  • No generated artifacts unintentionally added

renatobo and others added 7 commits February 22, 2026 18:32
Silence ultralytics startup banner and task-guessing warning by setting
YOLO_VERBOSE=false, logging level, and explicit task="detect". Update
make install-dev to include all scorer extras (clip, claude, yolo).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…utput improvements

- Add histogram correlation pass to dedup burst shots (same scene, shifted framing)
- Cache technical scores per image (keyed on mtime) to speed up re-runs
- Downsize images to 1024px/q75 before sending to Claude (reduces API cost)
- Show estimated Claude API cost before scoring (accounting for cached images)
- Add --rescore flag to force re-evaluation ignoring cache
- Add --work flag to put intermediate files in a separate directory
- Output three variants: full (original), hd (1920px), cropped (1080x1440 IG)
- Ducati brand priority in Claude/Ollama scoring prompts

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… parallelization

Burst detection:
- Three-tier matching: temporal+ORB (hist≥0.60), temporal (hist≥0.80+ORB≥0.25),
  non-temporal (hist≥0.92+ORB≥0.25)
- ORB feature verification prevents grouping different riders at same track position
- EXIF preserved during resize (orientation tag reset) for timestamp-based chaining
- Sharpness-first selection, then technical re-evaluation for top candidates

Performance:
- All CPU stages parallelized (ProcessPool for resize/dedup features, ThreadPool
  for tech scoring/burst re-eval/smart crop to avoid YOLO fork deadlocks)
- Claude API: adaptive concurrency (3→8), rate limit backoff, retry with exponential delay

Gallery:
- Built-in HTML gallery with detail panel (preview tabs, YOLO overlay, EXIF, scores, burst info)
- Recursive folder indexes with breadcrumb navigation
- Standalone regeneration via scripts/generate_gallery.py
- Crop icon badge for uncertain crops, burst count badge

Other:
- Default model changed to claude-haiku-4-5-20251001
- Brand bonus prompt for Ducati identification
- Images downsized to 1024px/q75 for Claude API to reduce cost
- Cost estimator uses resolved account context (fixes false cache misses)
- Three output variants: full (original), hd (1920px), cropped (1080x1440)
- --rescore and --work CLI flags

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@renatobo renatobo merged commit b59bd43 into main Mar 19, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant