• Registration is now open. It usually takes me a couple of hours or less to validate accouts. If you are coming over from VoD or DR and use the same user name I'll give you Adult access automatically. Anybody else I will contact by DM about Adult access.

Gemini's Reply to a GPT Negative Prompt - LOL!

Nyghtfall3D

Nyghtfall3D

Moderator
Staff member
I'm using the free version of Nano Banana on Gemini to shoot the shit with an idea for a cinematic sequence from a modern supernatural horror movie. It involves a 40yr old guidance counselor for a prestigious university who's leafing through her mail when she's attacked by the ghhost that haunts her house. One of the images NB returned looked stylized and cartoonish, so I had GPT give me a negative prompt tuned for NB.

This is what GPT it spit out:

Negative Prompt (Nano Banana – Ultra-Realism Lock):

cartoon, anime, illustration, painting, digital art, stylized, fantasy art, unreal engine look, CGI, video game render, doll-like, waxy skin, plastic skin, airbrushed skin, overly smooth skin, over-sharpened, oversaturated, neon colors, HDR halos, glowing outlines, fake rim light, studio glam lighting, beauty dish lighting, fashion editorial look

bad anatomy, distorted anatomy, malformed body, extra limbs, missing limbs, twisted spine, incorrect proportions, exaggerated curves, unnatural posture, stiff pose

face distortion, asymmetrical face, cross-eyed, lazy eye, dead eyes, glassy eyes, glowing eyes, exaggerated eye size, beauty filter, overdone makeup, smeared lipstick

bad hands, extra fingers, missing fingers, fused fingers, claw hands, deformed hands, incorrect grip

bad feet, extra toes, missing toes, fused toes, floating toes, broken ankles, bent heels, warped heels, incorrect heel height, thick straps, chunky straps, platform heels, closed-toe shoes

fabric clipping, clothing merging with skin, see-through clothing, latex shine, rubber texture, plastic fabric, incorrect satin sheen, stiff pleats, broken pleats

pantyhose errors, blotchy pantyhose, opaque stockings, fishnet texture, uneven nylon weave, tearing artifacts

hair errors, helmet hair, stiff hair, floating hair, chunky strands, plastic hair, inconsistent parting

environment errors, warped architecture, melted furniture, floating objects, incorrect reflections, mirror artifacts

low resolution, compression artifacts, jpeg noise, motion blur, depth-of-field errors, focus errors

oversexualized pose, pinup pose, exaggerated arch, unnatural leg spread, voyeur angle, fetish framing

text, watermark, logo, signature
Click to expand...

And this was part of Gemini's reply when I pasted the negative prompt in a fresh chat session but forgot to tell it what I wanted to create:

That is an exceptionally detailed negative prompt. You’ve essentially created a "reality filter" by systematically banning the most common artifacts found in AI-generated imagery.

By excluding things like "unreal engine look," "waxy skin," and "HDR halos," you are forcing the model to lean into raw, organic textures rather than the polished, synthetic aesthetic typical of base models.
Click to expand...
"reality filter"...! :ROFLMAO::ROFLMAO::ROFLMAO::ROFLMAO::ROFLMAO:

:: wheezing ::

Can't breathe!... :ROFLMAO::ROFLMAO::ROFLMAO::ROFLMAO:

:: more wheezing ::

Ow, my pancreas!... :ROFLMAO:

:: takes a few minutes to recover ::

On a more serious note, it went on to say:

Why this set works well:​

  • The "Uncanny Valley" Shield: By listing "doll-like," "plastic skin," and "glassy eyes," you're targeting the specific traits that make AI humans look "off."
  • Anatomical Rigor: Specifying "fused fingers" and "broken ankles" helps prevent the structural failures that often ruin an otherwise perfect shot.
  • Material Authenticity: The focus on "fabric clipping" and "stiff pleats" is a sophisticated touch—many users forget that clothing physics are often where realism falls apart.
Click to expand...

It then offered this tip:

A Quick Tip for "Nano Banana"​

Since you are using the Nano Banana model, it is highly responsive to lighting cues. While your negative prompt bans "studio glam lighting," you might get the best results by pairing this negative list with a positive prompt that specifies a "found" lighting source, such as:

  • Natural overcast light
  • Fluorescent overhead lighting
  • Harsh midday sun with sharp shadows
Click to expand...
Finally, it asked me if I wanted to generate an image using those constraints. So, I said yes, and here it is, a completely random image with no context:

coffee shop.png
 
Otto Von Herunterhängen said:
Technically with AI it would be like talking to all the voices in your head at the same time :ROFLMAO:
Click to expand...
Funny, I saw a video with Grok having a debate with Gemini, ChatGPT, Claude and Meta AI. Grok was the conservative voice while the rest were liberal. It was interesting, but scary as hell and pretty much what I expected after being to a bunch of townhalls and talking to real people.
 
Lol funny. It got me curious about the NB architecture. So it is a multimodal... Model, with the LLM portion directly interpreting the prompt, without a specific interpreting module. So that's why it can get pretty liberal with the creativity and styles, if stuff isn't specified exactly.

It's probably worth looking into json prompt formatting. LLMs are inherently good at using json and similar formats, and it can override most of the chattiness and randomness.

Try asking GPT to generate you a prompt in json format, or some blank templates where you can fill out the details. I've not tried NB myself because eff Google and all these censored dytopian models, but I bet it's gonna be more predictable. Or give it an example image, ask for a detailed description in json format and see what happens if you feed that right into NB.
 
You must log in or register to reply here.
Back
Top