Layout JSON vs plain prompt, 8 seeds each: when bounding boxes are worth it
SquareBat780 · reddit · 2026-10-22
A systematic comparison of prose-positioned vs layout JSON text rendering with MOSS-Image-SFT (bf16, ComfyUI v0.42.3, Euler 50 steps, CFG 3, SLG on layer 22), each design rendered 16 times across 8 seeds:
- 11-block price board: prose spelled everything correctly but placed only 67/88 blocks, with rows drifting downward in half the runs; JSON hit 88/88.
- Poster and label (4 blocks each): the two methods were within one block; spelling was perfect everywhere.
- Scoring required character-exact spelling and block centers landing inside target boxes with 5% slack.
Key gotcha: box coordinates are absolute pixels from the canvas center, so layouts break at other sizes — feed one width/height to both the Layout Prompt and Create Bounding Boxes nodes, or use the BBox1000 converter. Full built-in-nodes graph included. Verdict: the denser the text layout, the more JSON pays off.
More from Multimodal
- LiDAR scan + LLM workflow produces dimensionally accurate 3D room models — OpenAIDevs · 2026-10-22
- MERZraum: A Poetic 3D Project Built With AI Gaussian Splatting — Merzmensch · 2026-10-22
- ComfyUI Tutorial: Model Downloads, Folder Structure and Loaders Explained — pixaromadesign · 2026-10-22
- Side-by-side: Seedance 2.5 output vs free Muse looks nearly identical — sven_ai · 2026-10-22
- HeyGen ships a wave of updates: Hyperframes Studio, HeyGen Video and Pro Voice — toolstelegraph · 2026-10-22
- Combating MOSS Image blur with a mid-denoise latent sharpening trick — LMLocalizer · 2026-10-22