claude-video-image

claude-video-image — AI Image Generation for Video

Pre-Flight

1. Activate venv: source ~/.video-skill/bin/activate
2. Check free VRAM: nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits
3. Route to local or API based on VRAM availability and user preference

API Generation (Gemini 3 Pro Image) — Primary

Gemini 3 Pro Image (Nano Banana Pro) — best quality API generation, up to 4K resolution. Uses Google Search for factual accuracy. Best for photorealistic content and complex prompts.

source ~/.video-skill/bin/activate
python3 scripts/image_generate.py \
  --api gemini \
  --prompt "Professional marketing dashboard showing SEO metrics, clean modern UI" \
  --resolution 2K \
  --output thumbnail.png

Options:

• --api gemini — Use Gemini 3 Pro Image API (default API)
• --resolution 1K — Standard resolution ($0.13/image)
• --resolution 2K — High resolution ($0.13/image, default)
• --resolution 4K — Ultra resolution ($0.24/image)
• --aspect RATIO — Aspect ratio: 1:1, 16:9, 9:16, 4:3, 3:4, 4:5, 5:4, 21:9

Requires GOOGLE_API_KEY environment variable.

Local Generation (FLUX.2 klein 4B) — Free, Fastest

Best local option, free, sub-second generation. Requires 13GB VRAM (or 6GB with quantization).

source ~/.video-skill/bin/activate
python3 scripts/image_generate.py \
  --prompt "Cinematic wide shot of a futuristic city at sunset, 16:9 aspect ratio" \
  --width 1024 --height 576 \
  --output thumbnail.png

Options:

• --model flux — FLUX.2 klein 4B (default local, 13GB VRAM, <1 second)
• --model sd35medium — Stable Diffusion 3.5 Medium (6GB VRAM, ~5-20 seconds)
• --quantize int8 — Quantize FLUX to ~8GB VRAM (slight quality loss)
• --steps N — Inference steps (default: 4 for FLUX, 28 for SD)
• --seed N — Reproducible generation
• --batch N — Generate N variants (default: 1)

When to use each:

• Gemini 3 Pro Image: Best quality, complex prompts, factual content, up to 4K ($0.13-0.24)
• FLUX.2 klein: 13GB+ VRAM free, need speed, no API cost
• SD 3.5 Medium: <13GB VRAM free, or need different aesthetic
• OpenAI GPT Image 1: Need transparent PNG backgrounds

API Generation (OpenAI GPT Image 1) — For Transparent PNGs

Only option that natively generates transparent PNG backgrounds.

source ~/.video-skill/bin/activate
python3 scripts/image_generate.py \
  --api openai \
  --prompt "Professional podcast microphone icon on transparent background" \
  --transparent \
  --size 1024x1024 \
  --output overlay.png

Options:

• --api openai — Use OpenAI GPT Image 1 API
• --quality mini — GPT Image 1 Mini ($0.005/image, default)
• --quality medium — GPT Image 1.5 ($0.034/image)
• --quality high — GPT Image 1.5 high ($0.20/image)
• --transparent — Transparent background (PNG only)
• --size WxH — Output size: 1024x1024, 1536x1024, 1024x1536

Requires OPENAI_API_KEY environment variable.

Cost confirmation: Always show estimated cost and confirm before API generation.

Background Removal (rembg)

Strip background from any image to create transparent overlays:

source ~/.video-skill/bin/activate
python3 scripts/image_generate.py \
  --remove-bg input.png \
  --output transparent.png

Uses rembg with u2net_human_seg model (~2GB VRAM with GPU, or CPU fallback). Works with photos, AI-generated images, or any PNG/JPEG.

Batch mode:

python3 scripts/image_generate.py --remove-bg-dir ./images/ --output-dir ./transparent/

Video-Native Dimensions

Always generate at dimensions matching the target video format:

Use Case	Width	Height	Aspect
YouTube thumbnail	1280	720	16:9
Landscape video frame	1920	1080	16:9
Portrait/Shorts	1080	1920	9:16
Instagram square	1080	1080	1:1
Instagram portrait	1080	1350	4:5
Lower third graphic	1920	200	~10:1
Title card	1920	1080	16:9

For FLUX.2 klein, generate at 1024x576 (16:9) or 576x1024 (9:16) then upscale with FFmpeg:

ffmpeg -n -i generated_1024.png -vf "scale=1920:1080:flags=lanczos" output_1080p.png

Integration with Video Workflow

Generated images can be used as:

• Overlay input for the edit sub-skill (-filter_complex "overlay=X:Y")
• Background for green screen compositing
• Thumbnail alongside exported videos
• Title card for Remotion create sub-skill
• Ken Burns source for the screenshot sub-skill's animation pipeline

VRAM Management

• FLUX.2 klein requires 13GB — unload all other GPU models first
• SD 3.5 Medium requires 6GB — can coexist with light models (<5GB)
• rembg requires ~2GB — safe to run alongside medium models
• Script handles loading/unloading automatically via torch.cuda.empty_cache()

Safety Rules

1. Always confirm cost before API image generation
2. Run bash scripts/preflight.sh for output path validation
3. Never overwrite existing images without confirmation
4. Report generation time and VRAM usage in output

Reference

Load references/image-generation.md for model details, API parameters, and advanced options.