**Which models: **

• Gemini 3.1 Pro did better than GPT-5.6 on these tasks. It handled text and 3D room space more reliably. It still changes things it was not asked to change, so it needs tight change-and-keep prompting and a locked hero asset to stay in line.
• GPT-5.6 was better for REALISM but weaker for the style in the style guide. It tended to jump straight to generating without asking clarifying questions, and the output came out cheaper looking, with flat glow and generic sci-fi text. Usable, but needs more correction. So I'd use GPT if there's a very detailed change I want such as rendering a character in a glowy style -- GPT is more precise. But with the video gen, Gemini performs well with glow and dots. Model comparison with same prompt:

Install skill https://hyperframes.heygen.com/ for video generation.

Idealy, give reference photos and say exactly what to copy from the reference photo.

**What AI generation can do **

• Generate a strong still image in the AdaL style, given a good prompt and a reference image.
• Animate a still with a simple camera move: slow push-in, pull-back, gentle drift.
• Render text inside the frame, on the better models, if it comes out clean.
• Render a sense of 3D room or dimensional space, on the better models IF told to do so in the first prompt.
• Create atmosphere and light motion.

**What AI generation cannot reliably do **

• 100% hold a subject perfectly still across an edit. Tell it very EXPLICITLY what to change.
• Run long without drifting. Past about 5 to 6 seconds a generation degrades and faces melt. Generate 4 to 5 clean seconds and extend in post.
• Add grain and halftone the way we want. Covered below. This is a deliberate post step, not a model failure to fix.

**Why grain, halftone, and particles are done in post, not in AI **

• Control. In post you set the exact grain amount, the exact halftone density, the exact darkness, and you adjust them without regenerating anything. Ask the AI for grain and you get whatever it decides, and changing it means a whole new generation with a new random result. The texture is the thing that makes the work look like AdaL instead of a raw AI render, so you want to control it precisely, not gamble on it.
• Speed. Asking AI to add grain takes about two minutes and burns a generation. Doing it in CapCut takes seconds.
• Cost. Every regeneration costs money and time. Texture done in post costs nothing to redo. If you do not like it, you move a slider instead of paying for another render.
• Consistency. A post overlay applied the same way across every video keeps the whole set looking like one brand. AI-generated grain varies every time, so your videos would not match each other.

**The halftone overlay recipe ** There is a halftone overlay JPEG in the repo. To use it: Place the overlay on top of the footage. Set its blend mode to linear burn. Lower the opacity to taste. Higher opacity darkens the image, so grain, halftone, and darkness are all controlled from this one layer. Add particles AKA grain as an adjustment layer.

Because it is one layer with a blend mode, you can dial the whole texture and darkness of the video from a single control, and change it any time without touching the generation.

**The workflow this all points to ** Talk through the concept and ask clarifying questions: length, movement, 3D or 2D. Generate the hero asset, one still, and lock it. Animate the locked asset with a simple camera move, kept short. Add texture and any text in post, in CapCut.