Text To Image News
The latest Text To Image coverage curated for game developers.
A new study shows that fine-tuning open-source diffusion models on 1,000 captioned nuclear-energy images can materially improve text-to-image accuracy for specialized technical prompts. SDXL benefited the most, while SD-v3.5-Medium saw …
A new latent-identity tuning method aims to make text-to-image personalization far more precise for facial edits. It works inside a frozen encoder, finding semantic directions in latent tokens instead of …
A new benchmark, SciDraw-Bench, tries to measure whether AI can generate scientific figures that are actually usable, not just visually plausible. It focuses on 32 structured tasks across eight figure …
A new paper argues that prompt embeddings for text-to-image models already contain a usable semantic continuum, but the tokens are misaligned. By first rephrasing prompts into a shared structure and …
For graphics teams, the interesting bit is explicit control over defocus blur in text-to-image diffusion instead of hoping prompt wording fakes depth of field. The paper’s Bokeh Diffusion keeps scene …
A new arXiv paper tackles a practical problem for teams using text-to-image diffusion models: proving ownership even when a suspect model has had watermark signals damaged or stripped. Cert-LAS uses …
A new diffusion-transformer sampling trick aims to fix the “samey output” problem without the usual quality hit. By pushing repulsion into the model’s contextual attention space during the forward pass, …
A new approach to text-to-image generation, LGTM, allows developers to manipulate initial noise for better lighting control without extensive training. This innovation could significantly streamline workflows for graphics programmers and …
The introduction of Contrastive Noise Optimization marks a significant advancement in text-to-image generation, addressing the common issue of limited diversity in outputs. This method reshapes initial noise rather than optimizing …
A new open-weight diffusion model called Paris has been trained entirely through decentralized computation, without synchronized gradients or a central GPU cluster. It uses eight expert models and a lightweight …