Visuals, VFX and AI art: real first, AI only where no camera was

The client's real photos and clips are the film. Effects exist to make them feel alive and to show what no camera captured: a painted sea behind a real child, a road years from now seen from behind, a banyan tree of ancestors' names. Drawings must look like drawings, never like a fake memory.

The rules (non-negotiable)

1. No AI on real faces. No generative motion, edits, de-ageing or invented faces on real people, children above all. Image-to-video services may refuse photos of children anyway. tools/vfx/ai_gen.py refuses edit and image-to-video inputs that contain a detectable face.
2. Composite instead of generating over people. Cut the person out untouched (tools/vfx/cutout.py), generate or animate only the background, put the untouched cut-out back on top.
3. Style transfer keeps likeness or is rejected. An oil-portrait look on a real photo is allowed only with --allow-face-likeness, then compare faces before/after by eye and reject any change; lower the strength.
4. One style across every tool. Config style_suffix (a style sentence + five hex colours) is appended to every image and edit prompt automatically (motion prompts stay short and physical), so art from different models and days matches.
5. Quote gate and provenance. A paid call prints its quote first (ai_gen.py without --go exits 4), runs with --go --quoted-providers wavespeed only after the client's go, and stays under config.budget_inr; every call is logged to spend.md and out/ai_log.jsonl (prompt, model, inputs, seed, output). Copy each used asset's model and licence into sources.md.
6. Licences and labels. Check commercial terms per model and host before client work (some image-to-video routes have unclear commercial rights; some outputs carry a SynthID watermark). Label synthetic content where platform rules, ASCI or the IT Rules require (see ${CLAUDE_PLUGIN_ROOT}/docs/compliance-and-licensing.md sections 1, 3.4 and 4).

Decision ladder (stop at the first rung that works)

1. Frame it with a layout (prints, triptych, portrait, living) and the grade. Most shots end here.
2. Light and paper: LampLight, LightLeak, vignette, page drift, ink marks that draw on with the narration.
3. 2.5D living photo: python tools/fetch_models.py depth once, python tools/vfx/depth.py photo.jpg, then python tools/vfx/parallax.py film/plates.json (dolly ≤ 0.08, zoom ≤ 10 % over 5 s, faces still). Best on wide shots and landscapes; faces smear at depth edges.
4. Fix the source: straighten.py for framed portraits and documents shot at an angle; probe and fix rotation; use the best frame of a burst.
5. Composite: cutout.py (u2net_human_seg, Apache-2.0) + a painted or animated background.
6. AI still where no camera was: python tools/vfx/ai_gen.py image <model> out.png "<prompt>" --inr 8.
7. AI motion on a plate with no people: ai_gen.py video <model> out.mp4 "<motion>" --image plate.png, then retime.py (motion interpolation) to fill a longer shot: dreamier than speeding up a still. Look at it frame by frame; interpolation smears fast motion and thin lines.

Prompting (see references/ai-art-prompts.md)

Describe medium, composition, light and mood, never "photorealistic memory of <real person>". Ask for 2-3 variants, pick by eye in a still at film size, log the winner. For a recurring object (the family's real vehicle, house), lock its look from the proof photo in facts.md and say it in every prompt (model, colour, details): a wrong bike in an early test was another customer's bike.

Quality checks before an asset enters the film

• At film size in a still: edges of cut-outs (hair halo), seams, text artefacts, extra fingers, warped lines.
• Motion plates: watch at full speed and step through frames; no morphing, no flicker, no face.
• Matches the grade and palette; looks like art when it is art.
• Logged in out/ai_log.jsonl and sources.md with licence and cost.

Tool notes

• Host model routes and parameters change: read the model page before a run (<model path> in ai_gen.py is the host's route, e.g. an edit model or an image-to-video model).
• Large uploads can fail with SSL EOF on some hosts: downscale the input (max side ~2048) and retry.
• Open models fetched by tools/fetch_models.py are verified by md5: YuNet (MIT) and SFace (Apache-2.0) for faces, Depth Anything V2 Small (Apache-2.0) for depth; the Base/Large/Giant depth models are non-commercial. Never use bria-rmbg without a commercial licence.

References

• references/ai-art-prompts.md: prompt patterns, style suffix, plate specs, the composite recipe step by step
• ${CLAUDE_PLUGIN_ROOT}/docs/lessons-learned.md section F