---
name: upscale-start-frame
description: "Make or upscale a frame that will start an AI video (MiniMax H3 etc.): generate at the right size (GPT Image 2.5 Flare 2K for close-ups, 4K for wide shots, Nano Banana 2 4K) or upscale a 1K frame with Magnific. Part of the upscale meta-skill."
---

These creative references describe source experiments with external tools. Use your own supported account and tools; this download installs no automation. Settings and timings are experimental observations, not current FlashBoards availability.

# upscale-start-frame · the frame that starts a video

The video keeps the detail that was **drawn** into the start frame, not the file size (measured on MiniMax H3, same seed and prompt):
video from a 1K frame = 1.00× fine detail · from a native 4K frame 1.15× · from 1K + Magnific 4× 1.35× · from a resized frame 1.04× (noise).

## Choose

| Case | Tool · settings | Time |
| --- | --- | --- |
| 🏆 close-up, generate | GPT Image 2.5 **Flare** `openai/gpt-image-2.5/flare/edit` · quality **low** · image_size **2048×1152** | 19 s |
| 🏆 wide shot / full figure, generate | GPT Image 2.5 Flare · low · **3840×2160** | 35 s |
| 🏆 real 4K drawing | Nano Banana 2 `fal-ai/nano-banana-2/edit` · resolution 4K · seed | 43 s |
| already have 1K | Magnific `ultra-sublime` **2× `grainyAnalog`** (the best still) → then the video. Measured in video: 1K + 4× Portraits gave 1.35× detail (≈ native 4K); 2× Grainy not yet measured in video | not measured for 2× Grainy video |

## Rules

- Flare = Sunburst in look, faster (19 vs 24 s at 2K, 35 vs 84 s at 4K).
- GPT at 4K paints on close-ups — close-ups at 2K. GPT's pixel cap ≈ 8.3 MP (3840 on the long side only at 16:9).
- quality high added no visible detail in these generation tests — keep low.
- Hand the video model the full-size file; it downsizes by itself, the drawn detail survives.
- Never a plain resize.