From scaling to eFficiency LAws for viSual syntHesis
Sector: Cement • Location: France
Source: EU Funding & Tenders Portal
Current AI foundation models are trained on the entire Internet, consume enormous GPU resources, and follow the “scale is god” paradigm: the more data and compute, the better the results. Yet despite their scale, such models rarely produce the desired output on the first attempt, forcing users into repeated trial-and-error interactions. This inefficiency is most evident in conditional visual gener
- Status
- Prospective
- Sector
- Cement, Broadband, Telecommunications
- Location
- France
- Updated
- 19 Nov 2026
- Source
- EU Funding & Tenders Portal
ConstructChat
Ask ConstructChat about From scaling to eFficiency LAws for viSual syntHesis
Pick a question to edit it, or write your own.
First answer free · no account
Project Information FAQ
Project Information
Participants
| Not available for this record. |
Status
Original status | forthcoming |
Taiyo status | Prospective |
Taiyo last update | 2026-11-19 |
Contact
| Not available for this record. |
Description
Description | Current AI foundation models are trained on the entire Internet, consume enormous GPU resources, and follow the “scale is god” paradigm: the more data and compute, the better the results. Yet despite their scale, such models rarely produce the desired output on the first attempt, forcing users into repeated trial-and-error interactions. This inefficiency is most evident in conditional visual generation, where even lengthy, fine-grained text prompts rarely yield the desired image or video. FLASH proposes a paradigm shift: from scale to efficiency. We ask a fundamental question: Can we build models that reach high-quality, controllable visual outputs through the most efficient path, using the least data, compute, and interaction? We propose three efficiency resources with concrete methods for each: First, data efficiency: identifying how data quality, diversity, and type can replace raw volume. Second, data–compute efficiency: designing models whose capacity and training follow the data, using flow-matching with adaptive extra conditions. Third, interaction efficiency: leveraging multimodal user-model interaction by modelling abstract intermediates, such as sketch or layout; a user intent state; and targeted questions to reduce user–model turns. Ultimately, our goal is to uncover and learn efficiency laws that govern conditional visual generation and use them to build the next generation of multimodal foundation models, efficient by design. We will show impact on three applications, where efficiency is a game changer: (1) text-to-image generation, (2) multimodal generation and (3) robot manipulation, operating under strict constraints (memory, energy, latency). By replacing the race to scale with principled efficiency, FLASH will democratise conditional generation, promote sustainability, and enable deployment on constrained environments such as edge devices or robotics. We hope that FLASH will help create new models across diverse domains, accessible to all. |
Location
Country | France |
Source
Source | EU Funding & Tenders Portal |
URL |
More Details
Record type | Project |
Sector | Cement, Broadband, Telecommunications |
Article published date | 2026-10-15 |
Research
Free account
View all details
You are reading the public part. A free account opens the rest:
- Every detail on this project
- News and media coverage
- Technical research
Free account · no card · lands you back on this project
