8 measured lessons and 3 unsolved problems from building a 2h manhwa recap in ComfyUI
Sensitive-Wealth5801 · reddit · 2026-09-05
Building a 2-hour manhwa recap locally (RTX 4080 16GB, ComfyUI, Illustrious-XL + custom LoRA), the author shares hard-won measurements: expressions belong in the base pass, not FaceDetailer; low denoise is cumulative across passes; IPAdapterAdvanced leaks content (use PreciseStyleTransfer at 0.60); early endat breaks style; negations in positive prompts inject the thing; key attributes belong in the first 90 tokens; BiRefNet masks need hard thresholding; effects are compositing, not repainting. Queue-wait was 80% of wall time — use {"front": true} on POST /prompt. Three unsolved problems remain, including SetLatentNoiseMask failing to preserve faces at low denoise.
Related event: Local ComfyUI Experiment: Making a 2-Hour Manhwa Recap Video(2 posts)→
More from coding & agent
- Agent Plays Rimworld While Writing Its Own Mods and Taking Notes — jxnlco · 2026-09-05
- AI Agent Plays Rimworld, Writing Its Own Mods and Taking Notes as It Learns — jxnlco · 2026-09-05
- Perplexity Open-Sources Numbat to Detect and Forensically Analyze Rogue AI Agents — AravSrinivas · 2026-09-05
- theo praises GPT-6 Astra's async questions: models keep working without your answer — SIGKITTEN · 2026-09-05
- Dev burns $4k/day in tokens for $4k/month via 15 Codex + 5 Claude subs — zeeg · 2026-09-05
- Agents can do your research but can't yet turn it into a clear write-up — ZeroStateReflex · 2026-09-05