The First AI Art Generator That Can Spell: New FREE Open Source AI Art Generator

Text-to-image AI has long struggled with one basic requirement: producing readable, correctly spelled words. Mid-journey V4 can generate beautiful images, but its lettering often turns into gibberish or near-misses—prompting a years-long search for models that can actually spell. Google’s earlier “party” line of models showed incremental progress (from partial letters to occasional full words at very large parameter counts), but access was restricted, leaving most users stuck with models that can’t reliably write.

A new candidate aims to change that. A Google model described as a masked modeling system in discrete token space is positioned as a more efficient alternative to pixel-space diffusion approaches (including prior systems like DALL·E 2 and Muse-style diffusion). Instead of generating images pixel-by-pixel, it predicts randomly masked image tokens, with the input text first extracted from a large language model and then used to guide token prediction. Google claims state-of-the-art image generation performance—specifically including coherent spelling—while also reducing generation time through parallel decoding. The approach is also tied to “mask-free” and editing workflows: after an initial image is produced, the system can iteratively resample image tokens conditioned on a new text prompt to modify objects, swap elements (like changing the food next to a latte), or alter scene details (such as changing backgrounds from New York to Paris or adding hot air balloons).

The most immediately actionable news in the transcript comes from Deep Floyd AI, working in partnership with Stability AI. Deep Floyd announced an “if” model that produces sharply legible text in generated images—examples include clear, Photoshop-like lettering and consistent word shapes in multiple prompts. The announcement is framed around a playful poem (“if I were a model, I’d be open source”), and the model is presented as likely to be open source, with free variants available through websites similar to how Stable Diffusion is distributed. The transcript also highlights a range of themed generations—spray-painted phrases on walls, signage-style text, and character prompts—where the spelling is described as unusually accurate for text-to-image systems.

For users who want something to try right now, the transcript points to “Carlo,” a model that can produce text accurately in some cases. Carlo is available via a free iOS/Android app called B carrot discover, and it can also be accessed raw on Hugging Face. The results are described as prompt-dependent: simple words like “hello” can come out readable, though errors still happen (such as missing or extra letters). The model is labeled alpha, with a full release expected later.

Taken together, the message is clear: the next wave of text-to-image progress is shifting from “pretty pictures with messy text” toward systems that can reliably render readable words—either through Google’s token-based masked modeling approach, Deep Floyd’s open-source “if” push, or free early-access tools like Carlo.