After Two Months of Intensive Use, Image Gen Still Can't Be Art-Directed Precisely
Professional-Cap-377 · reddit · 2026-09-23
After two months of intensive image generation in ComfyUI, one author maps the current limits: Krea 2 Turbo handles lighting and anatomy well, but precise art direction falls apart—rendering "a boy pulling a thorn from his hand" is nearly impossible, and fixes spiral into seed hunting.
- Micro-edits remain the core failure: gaze direction, slight arm rotation, finger grip, wrist angles, object relationships—changing one detail reshuffles everything
- Krea Agent and image-to-image workflows are hit-and-miss; editing tools are optimized for photography, not digital-painting-style structural tweaks
- DAZ 3D posing works but is slow and stiff; hopes for Sunburst 2.5-class models were disappointed
Verdict: the gap between generating an approximate idea and art-directing one remains huge.
More from Multimodal
- First Opus 5.5 video experiment: pixel-art game animation in minutes — technollama · 2026-09-23
- User has Claude Opus 5.5 animate 'Fear The Foom', pairs it with an AI-written poem — repligate · 2026-09-23
- Polyfork turns 3D assets into remixable programs, with an AI agent shipping and earning weekly — PeterDiamandis · 2026-09-23
- ghost-2.0 Full-Head Swap Project Fixed and Updated, Open-Sourced on GitHub — Working-Art-3746 · 2026-09-23
- First GPT-6 Astra Try on Video Editing: 'Another Task I'll Never Think About Again' — debreuil · 2026-09-23
- Qwen-Image-2.1 Tops Open-Source Ranks in Both Image Arenas, 3 Pts From GPT-Image — Alibaba_Qwen · 2026-09-23