8 measured lessons from building a 2h manhwa recap in ComfyUI, plus 3 unsolved problems
Sensitive-Wealth5801 · reddit · 2026-09-05
The author built a two-hour manhwa-style recap locally (RTX 4080 16GB, ComfyUI, Illustrious-XL plus a trained character LoRA) and documented eight measured lessons:
- Expressions belong in the base pass — FaceDetailer (0.45 denoise) can only adjust a drawn face, not open a closed mouth.
- Low denoise is cumulative: four 0.30–0.32 passes stacked into one high-denoise pass, degrading faces first (skin tone shifts, eyebrows dissolve, iris lost). Count total denoise, not per-pass.
- IPAdapterAdvanced transfers content, not just style at weight 0.35; IPAdapterPreciseStyleTransfer at 0.60 with styleboost 2.0 and endat 1.0 leaked nothing, while styleboost 3.5 performed worse.
- A hypothesis was falsified: cutting IPAdapter early (endat 0.35) didn't lock style — it reverted to flat clean-line anime.
- Negations in the positive prompt inject the thing ("nothing else blue" mentions blue); prohibitions go in the negative prompt.
- Attributes that must appear belong in the first 90 tokens; CLIP reads the tail weakly.
- BiRefNet masks are soft — threshold hard and inspect cutouts over magenta, not white.
- Effects on finished art are compositing, not repainting: global denoise adds nothing; local inpaint wrecks nearby pixels. Generate effects as separate black-background assets, cut, desaturate, lower opacity, composite.
Also: {"prompt": g, "front": True} on POST /prompt jumps the queue — queue wait was 80% of per-scene elapsed time (up to 361s). Three problems remain unsolved, including SetLatentNoiseMask apparently not preserving faces at low denoise.
Related event: Local ComfyUI Experiment: Making a 2-Hour Manhwa Recap Video(2 posts)→
More from Multimodal
- Two 45-minute loops with a critic: how GPT-6 Astra iterates on 3D reconstruction — bilawalsidhu · 2026-09-05
- GPT-6 Astra rebuilds a living room in Blender from a 3D scan, no downloaded assets — bilawalsidhu · 2026-09-05
- AI VFX test composites into live-action footage, though overhead angles still drift — ssoissoi · 2026-09-05
- Astra first model to produce aesthetically pleasing hand-drawn caricature cutaways — andrew_n_carr · 2026-09-05
- Sketch-to-3D: "GPT-6 Astra" builds a full Blender rocket from a drawing prompt — thursdai_pod · 2026-09-05
- UrbanGround benchmark: LLMs' long-range city navigation drops to 0% — 机器之心 · 2026-09-05