Matthew

Just built my own version of Pinterest with @antigravity Gemini Argon, and it's insane!

é bom se previnir

Qwen image 2.1 vs Nano banana 2.1

I kind of like the details on the Nano Banana 2.1

nano banana 2 vs nano banana 2.1

An extension to extract images here, another tool to build prompts for pay there, and good old Gemini and I working with meta-prompts. BE LAZY

These past few days I've been experimenting with AI to create images that genuinely feel real. Check out these macro shots:

After using Gemini Pro in 2026 you only have 3 choices: - Keep paying while they increase token limits and burn your extra credits - Hit the wall every 5 hours with compute-based limits - Cancel the subscription and go back to GPT (if you're not rich enough to not care)

I already made a prompt for character sheets, so here’s a version just for extracting clothes. POSITIVE PROMPT Constraints: 16:9 horizontal, high-end editorial composite photo, split-layout flat lay composition of the garments and accessories from the reference image, studio setting. Composition: Left side (40% of frame): A 2x2 tight grid of four distinct close-up detail shots of individual items or fabric/print details (logo embroidery, stitching, fabric texture, tag details); clean isolated angles derived from the reference image. Right side (60% of frame): Two vertical flat lay shots side-by-side; First: full organized flat lay of all items arranged neatly together (top-down overhead view); Second: each item separated and individually spaced in a clean grid arrangement (exploded flat lay). Clear vertical gutters separating all panels. Environment: Professional minimalist photo studio; uniform off-white or light grey seamless backdrop; soft even overhead diffused lighting with subtle directional shadow to define fabric volume and depth; no harsh shadows. Items: All garments and accessories laid completely flat and wrinkle-free; items arranged with intentional spacing and alignment; no mannequin, no human body, no hands. Materials & textures: Preserve exact fabric colors, print graphics, embroidery, logo placement, textures and material finish exactly as shown in the reference image; faithful reproduction of all surface details. Primary focus (Realism): Realistic fabric micro-texture, visible cotton weave, accurate matte finish on black fabric; accurate orange print color with correct ink texture; no digital smoothing. Color & light: Neutral white balance 5500K; soft high-key lighting; medium contrast; true-to-life fabric tones matching the reference; no color grading or artificial glow. Camera behavior: Top-down flat lay perspective; 50mm–85mm lens aesthetic; deep depth of field; everything in sharp focus; subtle natural grain; zero over-sharpening. Overall aesthetic: Clean editorial flat lay; minimalist fashion catalog or merch lookbook feel; organized, precise, sophisticated. STRICT NEGATIVES human body, mannequin, hands, person, face, wrinkled fabric, messy arrangement, overlapping panels, fused panels, distorted items, altered colors, changed logo, modified graphics, plastic-looking texture, airbrushed, 3D render, cartoon, watermark, extra text, fake HDR, neon glow, oversaturated, dark backdrop, cluttered background, signature.

Nano Banana 2 vs GPT-Image-2 Dude, I don't know, I still prefer the Nano Banana 2. Prompt: { "subject": { "identity": { "reference_match": "exactly 100% consistent with the person in the uploaded image", "facial_features": "preserved, same expression, high-fidelity skin texture" }, "description": "A high-fashion editorial portrait of a woman sitting confidently on the front of a tank", "pose": "seated, legs slightly spread, relaxed posture, one hand resting between her legs, looking directly at the camera with a calm, moody gaze", "clothing": { "top": "crisp white button-down shirt tucked into trousers", "jacket": "deep burgundy/wine-colored structured wool blazer with a gold star-shaped brooch on the lapel", "bottom": "dark charcoal gray tailored trousers", "shoes": "pointed-toe metallic gold heels" } }, "environment": { "main_object": "T-90 Vladimir main battle tank", "object_details": "matte dark green military paint, visible reactive armor blocks (ERA), textured metallic surfaces, long main cannon barrel extending toward the left of the frame", "backdrop": "clean, minimalist blue-to-white gradient sky, evoking a studio-editorial atmosphere", "foreground": "A Dalmatian dog in the immediate foreground, intentionally out of focus (bokeh effect), adding a surreal and candid layer to the composition." ], }, "photography_and_style": { "composition": "Low angle, wide cinematic shot, center-weighted", "lighting": "Natural lighting, high-contrast, creating a moody and dramatic fashion atmosphere with deep shadows and bright highlights", "camera_settings": "Shot on 35mm lens, f/2.8 for subtle depth of field, sharp focus on the woman and the tank's armor", "color_grading": "Sophisticated editorial grading, rich burgundy tones, high dynamic range", "technical_specs": "8k resolution, ultra-detailed fabric micro-fibers, visible skin pores, authentic metal texture", "profile": "spc2tb2" } }

Everyone's obsessed with controlling every pixel, pose, and light source. I've been doing the opposite or so it seems. The goal was never less control. It's smarter control. Style rules. Strong reference assets. Then I let the model fill the rest. The results stop looking like AI tried hard. They start looking like someone directed something.