Problem Summary
Over the course of 10+ use cases, I’ve consistently encountered a major limitation in the current DALL·E image generation flow: the system fails to support true iterative correction or semantic consistency when refining images. This issue leads to wasted iterations, logical inconsistencies, and breaks in continuity, even when the prompt is perfectly explicit.
Typical Bug Example
When asking for a visualization of dimensional progression:
“0D → 1D → 2D → 3D → 4D”
(expected: point → segment → square → cube → tesseract)
DALL·E repeatedly skipped the 2D square, drawing instead a parallelepiped or cube, even after explicitly correcting the instruction to “draw a flat 2D square without perspective.”
Moreover, generating a “corrected version” ignores prior context and restarts from scratch, without incorporating the previous feedback.
Expected Feature Behavior
Iterative correction mode:
- Allow user to say: “Correct previous image by changing the 2D shape only”
- The model should reuse the prior image and context, not regenerate from zero.
- Logic-aware generation:
- Apply basic mathematical or semantic logic checks (e.g., dimensional consistency, shape progression).
- Flag obvious errors (e.g., skipping a dimension in a sequence).
- Reference continuity:
- Enable referential prompts: “Keep the same layout, just replace the 2D shape with a square.”
- Layered memory or object ID system (future-proof):
- Internally track elements (e.g., “Shape2 = square”) for consistent re-rendering and pinpointed corrections.
Why It Matters
For anyone doing technical illustrations, educational diagrams, or iterative visual design, this lack of control makes DALL·E nearly unusable for precise visual reasoning.
The quality of the image is not the issue — it’s the absence of continuity, iteration, and correction support.
Suggested Implementation
- Introduce a “revision prompt” mode with optional image state reference.
- Add a “freeze previous layers” option so only specified regions/objects are changed.
- Flag logical inconsistencies in the output when generating sequences or diagrams.
Would love to see this evolve — it’s probably one of the most limiting barriers for power users trying to build intelligent, modular visual workflows with DALL·E + ChatGPT.
