R&D FEEDBACK — IMAGES GENERATION WITH EXTREME CONSISTENCY AND EFFICIENCY No errors in even and odd numbers, solidified, and security and alignment disproportionate to human values.
Comic book generation needs to drastically evolve in its ability to maintain solid consistency across multiple panels, pages, and even and odd frame sets.
In my case, I provided a canonical reference image of the character Zarkon and a detailed script, but the generation still exhibited visual inconsistencies between panels. For comic books, this is a fundamental problem: the character needs to remain recognizably identical while only changing pose, expression, perspective, lighting, and action.
I would like a system capable of utilizing a kind of “structural visual memory” per character, automatically preserving:
• anatomy and proportions;
• face/helmet;
• armor and its parts;
• quantity and correct position of limbs;
• hands and fingers;
• weapons;
• colors and materials;
• accessories;
• symbols and luminous elements;
• Character scale;
• Spatial continuity;
• Direction of action;
• Relative position between characters;
• Lighting and setting when necessary.
A mechanism for temporal and spatial consistency between panels would also be extremely useful. Panel 2 should understand panel 1 as its previous state, panel 3 should understand panel 2, and so on. This would prevent each frame from being treated as an independent image.
For manga, a “Character Lock / Scene Lock / Prop Lock” mode would be excellent, allowing the user to mark a character, setting, or object as immutable and only allow explicitly requested changes.
Another important point is efficiency. The research should aim for a significant reduction in:
• Computational cost;
• Latency;
• Token consumption;
• Memory required;
• Electrical energy;
• Cooling requirements;
• Waste from unsuccessful generations.
The long-term goal should be to make generative creation much cheaper, more efficient, and more accessible, allowing ordinary people, students, artists, independent developers, and creators to experiment with AI without each attempt incurring a high cost.
It would also be interesting to research highly efficient hardware and software architectures, including specialized accelerators, model compression, advanced quantization, sparsity, visual state reuse, caching, and generation techniques that avoid recalculating information that has already been established.
My long-term vision is an extremely general, efficient, and safe AI, capable of performing complex creation, programming, research, and engineering tasks on hardware much more efficiently than current systems. A future generation of models—for example, something along the lines of a hypothetical GPT-7—could be used as a research platform to develop even more efficient models and chips.
The goal should not simply be “bigger models,” but more intelligence per unit of computation, energy, and cost.
Ideally, as technological costs decrease, there should be free or extremely affordable options for creation and entertainment, allowing as many people as possible to explore creativity, education, programming, art, games, manga, and research.
Ultimate vision: drastically reduce the marginal cost of artificial intelligence while simultaneously increasing its capacity, security, efficiency, consistency, and public accessibility.
In particular, I would like to see comic book generation evolve from “generating multiple similar images” to “rendering a visually continuous story.” This difference would be revolutionary.
And I would add a specific requirement for your case:
COMIC BOOK CONSISTENCY TEST: provide a 20–50 panel character + script reference and automatically measure whether the same character remains consistent across all panels, including face, armor, hands, weapons, proportions, and visual identity. The system should detect errors before delivering the page to the user and regenerate only the problematic panel, instead of wasting the entire generation.