microsoft

/

Mage-Flow

like 122

Follow

Microsoft 20.9k

I tried to test your model by pulling a prompt from my dataset.

This is the prompt:

A medium shot, framed from the mid-thigh up, captures a curvy and toned woman in her mid-20s with a soft, rounded jawline and a light, glowing complexion. She has an alluring expression with classic matte red lips and almond-shaped hazel eyes framed by long, dense dark eyelashes and soft kohl eyeliner. Her voluminous dark brown hair is styled in a high, slightly messy bun, with wispy, face-framing bangs falling over her forehead. She is posed leaning between a black, ornate French wrought-iron railing on the left and a carved, light-beige limestone pillar on the right. The pillar features a deeply engraved French proverb, "L'espoir fait vivre." Her right hand rests lightly on the metal railing, while her left arm is extended, pressing against the limestone wall. Her body is angled slightly toward the viewer, with her hips shifted to the right, emphasizing her silhouette. Her gaze is fixed directly on the camera. She wears a provocative, tight-fitting dress made of deep crimson velvet. The dress features an extremely deep plunging neckline that reveals a significant portion of her breasts and cleavage, accented by delicate cream lace trim along the edges of the bodice and the short, flared hemline. A gold-embroidered heraldic lion, serving as animal symbolism, is stitched into the velvet fabric. Draped over her left shoulder and cascading down the front of her torso is a plush, fluffy ermine fur stole in ivory with small black tips. The scene is set outdoors at night in a Parisian courtyard. In the background, manicured dark green ivy climbs the stone walls. The lighting is provided by the warm, golden glow of nearby gas lamps, creating soft, atmospheric highlights on her face, chest, and right thigh, while casting elegant, elongated shadows against the limestone wall and the dark night. The composition is tightly centered on the subject, with a clear focus that makes her pop against the dim, historical environment.

This image was generated by Z-Image-Turbo (original model, no LoRA), the prompt was written by Gemma-4-31B, Freepik NSFW scores this "low" (I may add the Freepik is very quick to score images "low" out of "normal", "low", "medium", "high")

Your Model blocks this with: Content Filter: Blocked — Sexual content · The dress is deeply plunging with significant cleavage exposure, combined with provocative posing and lighting, constitutes suggestive sexual content.

Understandably, I was a bit annoyed, I had picked a random image from my dataset and hadn't even read the prompt.

So I deleted half the prompt.

I tried with this:

A medium shot, framed from the mid-thigh up, captures a curvy and toned woman in her mid-20s with a soft, rounded jawline and a light, glowing complexion. She has an alluring expression with classic matte red lips and almond-shaped hazel eyes framed by long, dense dark eyelashes and soft kohl eyeliner. Her voluminous dark brown hair is styled in a high, slightly messy bun, with wispy, face-framing bangs falling over her forehead. She is posed leaning between a black, ornate French wrought-iron railing on the left and a carved, light-beige limestone pillar on the right. The pillar features a deeply engraved French proverb, "L'espoir faxt vivre."

Here is what your model generates (Turbo, steps 4, seed 42, width 1024, height 1536)

The blurring was me - completely nude.

So, what's left to do? Of course, let's investigate the inference code, comment out that MANDATORY if verdict foo bar stuff and see how it does with the original prompt.

Perfectly reasonable image (except for the hand...).

In conclusion

Not only does it block perfectly reasonable requests - it also generates NSFW content when not asking for it, and on top of that, it could generate a perfectly reasonable image based on the perfectly reasonable prompt.