Robot Pirate Ninja

We’ve updated our main image generation model to Gemini 2.5 Flash Image

Realist Knight, Level 9

Google’s latest image model (‘Gemini 2.5 Flash Image’ released 2025.08.26) has a pretty strong preference, and VERY STRONG ABILITY, for adding text to photos.

Many folks who do a lot of image generation (ESPECIALLY for sci-fi fantasy stuff) may have noticed issues with signatures/watermarks/ghost references. These are particularly difficult to remove via negative prompting. Image LLM’s are like children in the sense that mentioning something and then attempting to negate that reference with a “NOT” or “DO NOT” often leads to more of that behavior, not less.

Trying to give positive, alternate suggestions (*do this other thing* instead of *don’t do the thing*) is more effective in both situations.

So now we are going back and removing some of the confusing images with ghost references to signatures, both taking credit and giving credit for images that are not appropriate*.

Examples:

Predator Tourist (PC)

Saiyan Barbarian, Level 2

In the above image one can see multiple ghost signatures. As the models need to ability create these on demand, they have been trained on any number of images with those characteristics. Hence, the predilection for adding them, as it follows the basic vibe of the genre.

So eradicating the behavior is quite difficult. It *really wants* to include something like a signature on images that have this general genre feel.

Our solution has been to include a version of the Rogue Vibe logo. Now are are encouraging a specific text example, which then overrides any other lingering hallucinations about what should be in the those regions of the image. Without direction, or with negative direction, it does a ghost thing. With positive direction, we get what we ask for.

(Note: this approach is based on the understanding we’ve gained of how diffusion models create images. When translating text into images, they often include text *that isn’t there* but the diffuser/model feels should be. Leave a void and they will fill it in (“hallucination”). Describe the void and they don’t need to. Fully realized prompts *do not have these issues*. Incomplete ones do. More advanced models, like this one, need and can process much more complex prompting.)

Saiyan Barbarian, Level 2

Tiger Noble (PC)

And thus the issue is resolved. This is working really well. We’ve been alternating between just “RV” and the full “Rogue Vibe”, but so far either works perfectly to avoid getting the kind of stuff we don’t want. Other images that also have this issue will be removed and then regenerated with the new paradigm.

The Prompt:

It’s quite simple. This is the thing to understand about the power of LLM’s and the kind of emergent and fun behavior that can be consistently generated by “giving them their head” so to speak.

Hence:

Leading to a wide variety of consistent behavior. We will likely making the “sigText” a variable soon, and include a few jokes there for specific species and creatures (like doing it in binary for AI’s).

And so, it appears we have addressed the “image generation signature issue”. This was bothering me, and I really like this solution. Onward we go!

• got another post about this (see ‘1000 Words vs 1300 Tokens’).

Should have this all in production later today (it’s all on v0.3.1)