NEW Text to Image AI "Simulacrabot" Compares to DALL-E 2 & is OPEN SOURCE!

A new open-access text-to-image bot called “Simulacrabot” is drawing comparisons to major paid systems by pairing Stable Diffusion with a highly curated dataset of synthetic images and human aesthetic ratings. The core advantage isn’t just raw generation—it’s the way the model is tuned using “Simulacra Static Captions,” a dataset built from more than 238,000 AI-generated images. Those images come with caption-and-rating triplets derived from over 40,000 user-submitted prompts, then scored for aesthetic quality, producing a royalty-free resource with more than 176,000 ratings that can also be reused for downstream research and training.

Simulacrabot is accessed through Discord and operates under early-access rules: no NSFW content, no copyrighted material (such as generating Disney princesses), no hateful content, and no personal information. Despite those guardrails, the examples shown are striking for how closely the outputs follow prompt style and for how quickly the system returns results. The model produces coherent, visually pleasing images across multiple categories—origami “Godzilla” destroying origami Tokyo, near-photoreal portraits (including Albert Einstein), stylized oil paintings that still read clearly as a single scene, and character concepts like an orangutan described with a trending ArtStation prompt.

The most direct comparisons come from side-by-side prompting against DALL·E 2. In one test, both systems generate an aesthetically strong image, but the Simulacrabot output is described as missing only a small amount of “coherency” relative to DALL·E 2. The transcript repeatedly frames this as a competitive gap that could narrow further as open-source pipelines improve—especially with free upscalers that can raise resolution. Another test targets photorealism: a “nature photograph of a frog” looks convincing enough to be mistaken for a real photo, though the frog isn’t perfectly centered. Re-running the same prompt yields a slightly different composition, reinforcing that the model is responsive but not always perfectly aligned.

Simulacrabot also appears strong at faces, a notoriously difficult area for text-to-image models. Examples include a photorealistic red-bearded man eating ice cream and multiple “Grand Theft Auto 5 cover art” style celebrity depictions, including Jeff Bezos, Mark Zuckerberg, and Elon Musk. The outputs are presented as accurate enough to be recognizable, with the “stitched art” look matching the GTA cover aesthetic.

Overall, the takeaway is that open-source text-to-image systems—when trained or tuned with large, rating-based aesthetic datasets—can reach a level of visual quality that rivals mainstream paid tools. That puts pressure on pricing and performance assumptions for closed models like DALL·E 2, particularly as iteration speed and community tooling accelerate.