A GenAI comic pipeline cut iteration passes from 20+ to under 5, then lost on reception

Richard Rondeau has published a build log for Veridian Resonance, a bilingual digital comic he treated as a production pipeline rather than an art project, and the process numbers in it are specific enough to argue with ^[1]. He also reports that the output landed flat with viewers, and that reception, not a technical wall, is what is pushing him back toward visible manual drawing ^[3].

Start with scope. The project shipped 100 panels generated and refined across 24 finished pages of published work ^[2], roughly four panels per page ^[20]. It did not begin as illustration: the world came out of a homebrew tabletop campaign, managed as interconnected markdown nodes in Obsidian and formatted through Homebrewery ^[4], with the stated engineering problem being how to scale dense relational text, lore, mechanics and rigid character definitions, into a 28-page visual production ^[5].

The style choice was made on production criteria rather than taste. Rondeau says he ran structured A/B testing across aesthetic checkpoints scored on panel consistency, manual rework required, and reproducibility at volume ^[6]. Grim Dark and Marvel looks were rejected because they typically needed 20 to 30 iterations per panel before post-processing to fix hatching and lighting inconsistencies ^[7]. The replacement was a strict cel-shaded formula, bold black outlines, flat colour, minimal ink shading, limited palette, with structural reference to Vox Machina and Avatar: The Last Airbender ^[8].

The headline efficiency claim comes from restriction, not better prompting. Negative prompts were tuned to block the model's pull toward generic 3D hyper-realism, producing a baseline stable enough to colour-grade against predefined palettes in Krita ^[9]. Combined with reference locking, that cut iteration cycles per character asset from 20-plus passes to under 5 ^[10], which is at least a fourfold reduction on the low end of the stated range ^[11].

Consistency across a sequence was the harder problem, and the first approach failed. Local models fine-tuned with LoRA held the art style but made character consistency across panels nearly impossible, a failure mode Rondeau says he later saw again in newer Midjourney versions ^[12]. The working system used Nano Banana 2 with character sheets and templated prompts to hold the lead character stable across the run ^[13], with two sheets per character, one face and one full body, which he found preserved detail better than a single reference ^[14]. Non-standard designs needed explicit structural commands: Krix required scale instructions to render as a short goblin instead of drifting back to a kobold archetype ^[15].

Around the image work, custom scripts removed an estimated 15 hours of manual file management, naming and formatting ^[16], and the site shipped as a bilingual EN/FR static build with no duplicated assets across language variants, swapped at build time ^[17].

Then the part that matters for anyone costing out a similar pipeline. Rondeau says he recognised a quality ceiling in the pure-AI approach and made a deliberate architectural pivot for Chapter 2 rather than pushing more prompts at the same problem ^[18]. Reception on personal and niche channels stayed muted, with much of the work read by viewers as "AI slop" regardless of the craft behind it, and that is what informed the shift toward more visible manual drawing, framed as a credibility decision as much as a quality one ^[3].

Worth watching: whether Chapter 2's hand-drawn content moves reception at all, whether the automation and reference discipline survive the pivot or get abandoned with the generation step, and whether the remaining four pages between 24 published and the 28-page target close under the new method ^[19].