Provenance
The prompts
Some imagery on this site was created with AI assistance. In the interest of transparency, here is every prompt and setting that made the descent, verbatim and unedited. The same records are mirrored as text files in the/prompts folder of the repository.
1. The descent, draft takes
Higgsfield MCP, generate_video
- Model
- Kling 3.0 Turbo (kling3_0_turbo)
- Resolution
- 720p (draft)
- Duration
- 10 seconds
- Aspect ratio
- 16:9
- Count
- 3 takes (compared on mid and final frames)
- Cost
- 15 credits for the three takes
- Preset
- Declined the suggested IN THE DARK preset, generated literally
Prompt, verbatim Slow super dolly-in descending straight down through deep abyssal ocean water, near-black background gradually revealing drifting clouds of bioluminescent plankton and glowing cyan-teal particles, soft volumetric shafts of blue-green light filtering down from far above, tiny sparks of light pulsing and blooming softly in the darkness, a faint rare magenta-pink bioluminescent glow deep in the frame, the camera creeps forward and downward at a single steady almost imperceptible pace, no hard cuts, one continuous slow tempo, shallow depth of field with gentle cinematic bokeh on the nearest particles, fine film grain, deep cinematic teal-and-black color grade, 24fps film look, ultra-detailed particle simulation, settles into a subtle breathing-like loop with position and lighting unchanged from the first frame
Outcome Three takes generated. Take 1 was chosen: the cleanest arc, a dark abyssal descent through volumetric light shafts that ends on a single, bright, centred magenta bloom. Take 2 (richer particle density) was the runner-up. Because scroll drives playback, a one-way descent that blooms at the end beats a seamless loop.
2. Upscale the winner
Higgsfield MCP, upscale_video
- Model
- ByteDance video upscale (bytedance_video_upscale)
- Source
- Take 1, 1280x720
- Target
- 2K (2560x1440)
- Preset
- aigc (tuned for AI-generated footage)
- FPS
- 24
Outcome Sharpened the drifting particles, the light shafts and the bloom before frame extraction, so the 1920x1080 desktop frames are crisp rather than upscaled beyond native.
3. Extract WebP frames
scripts/extract_frames.py (ffmpeg + libwebp)
- Model
- Local pipeline, no model
- Frames
- 140 (evenly spaced from the 2K source)
- Quality
- WebP q88, lanczos scaling
- Desktop set
- 1920x1080, 140 frames, 2.01 MB total
- Mobile set
- 960x540, 140 frames, 0.75 MB total
- Scroll height
- 450vh sticky descent
Outcome Both framesets sit well under the 10 MB desktop / 5 MB mobile budget. The dark water compresses beautifully, which bought headroom to raise quality and frame count for a smoother scrub. Critical frames load first to protect the LCP.
Want the full build method rather than just the prompts? Readhow it was built.