Fixing broken text in AI slides: generate the design, strip the text, then let AI refill it

algo_diver · x · 2026-09-08

A practical workflow for making PPTs with image-generation models without the usual garbled-text problem:

• Human drafts the outline first: prepare the message and copy for each slide (text alone suffices), so AI executes rather than decides;
• Generate many design variants: with (almost) the same prompt covering style, tone and information density, pull 10-100 slightly different drafts;
• Ask the model to remove all text from the chosen design, yielding a clean layout;
• Have the AI insert the finalized text back onto the text-free image, handling font, spacing and fitting details.

The core trick: split visual generation from text layout to sidestep image models' weak text rendering.

Original post →

More from Multimodal

• GPT-6 Astra rebuilds a classroom from one photo using Blender CLI and MCP — sidahuj · 2026-09-08
• Dev compares Blender MCP vs computer use: one-shot results favor MCP — sidahuj · 2026-09-08
• Building a castle with GPT-6 Astra and Blender MCP took just one hour — sidahuj · 2026-09-08
• Controlling AI speech emotions with microexpressions, tags and context — a-ijoe · 2026-09-08
• Local MiniMax H3 on RTX 5090: 15s 2K video in ~9 minutes via a 3-stage pipeline — Hailuo_AI · 2026-09-08
• Spectrogram-to-Sound AI Meets Bioacoustics for Deeper Bird Vocalization Research — CatAstro_Piyush · 2026-09-08

Multimodal channel →