edenfunf / blender-product-film

Blender product film

Everything is driven by JSON shot specs → scripts/pfilm.py (runs inside Blender) → PNG frames → scripts/finish.py (edit, titles, audio).
Blender 4.2+（BLENDER 環境變數或在 PATH 上；建議有 OptiX/CUDA 的 GPU）。選配 Real-ESRGAN（REALESRGAN 環境變數指到 realesrgan-ncnn-vulkan）。

Pipeline

build/
  product/        prep_product.py output (from the user's photo)
  shots/S1.json   one spec per shot (format: top of scripts/pfilm.py)
  renders/…       frames
  edit.json       the cut: order, transitions, titles, music, SFX (format: top of scripts/finish.py)

1. Product. User photo (front view, plain background) →
   python scripts/prep_product.py <photo> --out build/product --height <real height m>
   Small images are AI-upscaled 4× first. Open build/product/check.png and confirm the green silhouette hugs the product and the
   magenta print mask covers the label/cap and not the clear body. No photo → procedural "type": "bottle" (see pfilm.py).

2. Shot specs. One JSON per shot in build/shots/, written from the plan's shot table (camera language below).

3. Look-dev stills (fast, do this first, iterate here):
   python scripts/render_shots.py build --quality final --stills auto (start/middle/end of each shot) → out/check/<id>_stills.jpg

4. Motion preview of every shot + a rough cut compared to the reference:
   python scripts/render_shots.py build --quality preview --cut → out/check/<id>_preview.jpg, out/preview_cut.mp4, compare sheets

5. Compare with the reference (see Review) and fix; repeat 3–5 until every shot passes.
6. Final: python scripts/render_shots.py build --quality final (≈1.5–4 s/frame at 1080p) → python scripts/finish.py build/edit.json.

Camera language (use these; name them in the plan)

Every shot has a move, a subject, a fill and a motivation.
Speed: match the reference, measured. Premium films move slowly — the reference's shot_details[].camera.pace is usually slow/medium.
The ranges below are the total travel for a ~3 s shot at a slow pace; macro shots need far less travel than full-product shots
(on a small region, fill 0.95→0.85 already reads as a 50% zoom). Always: preview → --cut compare → scale with camera.speed_scale
(ratio reference speed ÷ our speed, floor ≈ 0.2 so nothing goes dead) → re-preview, before the final render.

Rules of the look (what makes it read "premium"):

• Black stage, low key. set.backglow gives a soft coloured pool on a dark backdrop; key light low (≈5), rim strips do the modelling,
  one moving sweep light per hero shot. Never a flat bright background.

• The product fills the frame. Hero ≥ 0.7 of frame height; details via region + fill ≥ 0.8. A small product in a big black frame is the #1 failure.
• Every shot moves (camera or product or light), slow and eased — no cut to a dead still.

• Open with a reveal: shot 1 starts dark and the light finds the product (lights.reveal: [0, 1.5]), centred or on a third with
  purpose. It is the most striking image of the film — never a static, half-empty frame.

• Empty frame has a job. More than half the frame black and empty is only allowed when a title lives there. Tall thin products
  (bottles) cover little area even at full height: crop tighter (cap-to-label), go lower and closer, or put the title in the empty side.
  pfilm prints PFILM_WARN (and meta.json) when the product is small or its texture too soft for the close-up — treat warnings as defects.

• Close-up rims: pfilm switches to a 4 mm strip at equal radiance on close-ups (a fine edge line). Don't force closeup_rim up:
  a thick continuous white outline reads as a cut-out sticker.

• Titles: headline cards 0.06–0.09 of frame height, supers next to the product 0.04–0.05, closing brand line ≥ 0.05.
  Never three text beats in a row without a product image between them unless the reference does exactly that.

• Depth of field by depth, not f-number: "dof": "shallow|medium|deep" computes the f-stop for the camera distance
  (macro at f/3.5 leaves ~2 mm sharp — unusable). Focus sits on the product's front surface automatically.

• Titles are 2D, in negative space, never over the product: short (3–8 characters/words), grey + white two-tone, light_up or rise reveal.
  Title-only frames (black + text) ≤ 15% of runtime — a product film is pictures first.

• Macro on bubbles / droplets needs a bright area behind the product (clear things only show what they refract): bring the backdrop
  close (set.backdrop_dist ≈ 0.55) and make the back spot small and bright (backglow_spread ≈ 30, strength 9–16, backglow_height ≈ 0.25).

• Light sweep energy ≈ 6–10. A bright strip passing between lens and product washes the label and cap to white.

• Atmosphere (pfilm set): ground_fog = low rolling haze at the base only (density ≈ 0.3, height ≈ 0.08 m) for "mist / cool" beats —
  never a whole-frame grey wash. caustics = moving water-light lines on the backdrop for "ocean / water" beats; best on a mid-blue
  backdrop (backdrop_color), keep it subtle (default strength 0.35) and let depth of field soften it.

• Rising bubbles are product.bubbles (inside the water); condensation is product.droplets (outside). A shot that promises bubbles must show bubbles.

• Clear liquids need light BEHIND them. In a dark field, water reads as black ink unless something bright sits behind it:
  keep a soft bright zone behind the product in every dark shot (set.backglow strong and low behind the body, or a glow card),
  so the liquid refracts light and reads clear. "Is the water clear?" is the first question for any drink film.

• Titles say something. Repeating the product name is not a message: each text beat is one claim (taken from the pack or the
  user) followed by the picture that proves it. The name appears at the open and the close.

• Don't repeat a detail: the same close-up subject (e.g. the cap) at most twice per film; vary subject, scale and angle between
  neighbouring shots. Keep the product's colours identical across shots — change light level, not hue.

• The last shot is a hero: full product or shoulder-and-label, fill ≥ 0.7, a final light sweep across the name, and ≥ 1 s of stable
  frame before the fade.

• Proof shots: turn each claim into a picture of the product (droplets = cold, bubbles rising = fresh, macro label = the name, orbit = the shape).
  Only claims that are printed on the product or given by the user; never invent specs for a real brand.

• Cuts on motion: the next shot continues the direction of the last (orbit right → next move right), or dips through black at section breaks.

Review (mandatory before the final render)

For every shot, put the reference shot it maps to next to ours (video-clone scripts/compare.py), open it, and score 1–5:
lighting (low key, rim, highlights) · material realism · camera motion (type and speed match the reference) · composition / fill ·
typography · pacing. Any score < 4 → fix the spec and re-render stills/preview. Write the scores to out/check/review.md.
Also open one full-res still per shot and look for: blown highlights, washed-out label, visible light panels, clipping into the product,
text over the product, flat black frames.

Known pitfalls

• EEVEE's screen-space refraction breaks clear bottles (black bands) → preview and final both use Cycles (default).
• Big emissive cards behind the product flood the frame → use set.backglow (backdrop + spot), not an emissive wall.
• Glossy floors mirror the softboxes → lights are light-linked to the product (automatic).
• Blender renders CJK blank from Noto variable fonts → 3D text uses Microsoft JhengHei; but titles belong in finish.py anyway.
• Web product photos are tiny (≈400 px) → prep upscales with Real-ESRGAN; still, avoid extreme label macros (fill > 0.95 on text).
• Baked highlights in the product photo are removed from the print mask (thin white outlines are kept, semi-opaque).

• Ghost / doubled label text → check print.png's alpha first, not the lights. Upscaled small CJK text leaves speckles between strokes.
  Best fix for a user's own product: clear the name area in print.png and re-typeset the printed name in place (same text, colour,
  position; 2× resolution), then update product.json (m_per_px, bottom_px) — see a project's build/relabel.py for a worked example.

• Photo relief (bump from the photo) is off by default: it turns JPEG noise into embossed grain.
• Hue-matched print (a blue stripe on a blue-tinted bottle) gets cut out as "clear body" → patch those holes in print.png.