رفتن به محتوا
newblender

09 — Il Baracchino: how an agent could make it

این محتوا هنوز به زبان شما در دسترس نیست.
Original Megadrago (Palermo; creators and directors Nicolò Cuccì and Salvo Di Paola) with Lucky Red, 2025, an adult animated TV series for Prime Video · blender.org user story
Agent-readiness today Partial. The series’ signature tricks are unusually agent-friendly because they are rigs and data, not hand drawing: UV Warp tile swaps for cut-outs, stepped keys with holds, a GN-built mouth, EEVEE, and VSE editorial. What needs tooling or a human: Rigify generation and fitting (operator- and context-driven), facial acting, comedy timing, and the art of the puppets.
Difficulty 4. Six episodes, about 100 minutes, made by about 20 people. A single 30-second scene is a credible stretch goal.
First slice A 10-second, black-and-white “comedy club” beat: a box-headed cut-out puppet whose face swaps between 5 expression tiles through a UV Warp modifier, a rubber-hose arm on bendy bones, everything keyed stepped on twos/threes, rendered in EEVEE and cut into 3 shots in the VSE.

1. What the original actually is

Section titled “1. What the original actually is”
  • Series: 6 episodes, about 100 minutes in total. Italy’s first adult animated series made in Blender, premiered on Prime Video on 3 June 2025 and presented at Blender Conference 2025. It follows an art director trying to save a failing comedy club.
  • Look: “primarily black and white”, with one episode in colour (user story). The series mixes 3D, stop-motion-style puppetry, paper cut-outs and 2D.
  • Characters (named in the source):
    • Claudia (protagonist): Rigify with bendy bones for rubber-hose arms, and shape keys plus bones for the face.
    • Leonardo da Vinci: “five texture tiles mapped onto simple box geometry”, driven by UV Warp.
  • Output rate: “40 seconds per week” at peak.
  • Team and stack: about 20 people, Blender 3.6 LTS, EEVEE, the VSE for editorial, Kitsu (tasks), Flamenco (render management), SVN (versioning).

2. How the humans made it

Section titled “2. How the humans made it”
Stage Fact (user story) Interpretation
S04 GN Mouth = “curve connected to bones, with the gradient achieved procedurally through geometry nodes” The mouth shape is driven by bones, and its look is computed rather than painted per frame
S05 Rigging UV Warp is “foundation for multiple characters, allowing us to create paper cutout-style animations”; Rigify plus bendy bones; shape keys combined with “traditional rigging” Cut-out = swap texture regions instead of deforming geometry
S07 Animation “stepped mode with holds” only, no spline animation; timing on “twos, threes, fours” Fakes stop-motion cadence. Animators choose the step per action.
S12 GP Grease Pencil for “3D previews with 2D animated characters” (previz) GP as a sketching layer, not the final look
S09 EEVEE NPR; layout set lighting intent before the stop-motion integration Fast renders made 40 s/week feasible
S11 Blender VSE for editing Editorial stays inside Blender

3. How I would make it: the agent-native plan

Section titled “3. How I would make it: the agent-native plan”
Stage Agent approach Inputs (free) Tooling
S02/S03 Expression atlas generated as pixels: a 5×1 image atlas (eyes and mouth drawn as filled ellipses and arcs) written via image.pixels, black ink on paper colour, plus a paper-grain noise. Or use human-drawn tiles if the owner supplies them. CC0 paper textures (Poly Haven / ambientCG) bridge
S05 cut-out UV Warp as the expression switch: a box head with UVs covering one tile; UVWarpModifier offset U = tile_index / 5. An integer “expression” custom property drives the offset through a driver. — modifiers, drivers
S05 body Armature as data. Arms as one bone each with bbone_segments = 8 and ease in/out keyed, giving the rubber hose. Rigify: enable the add-on, build a meta-rig and generate (pose.rigify_generate; context needs are untested) for the full version only. — bridge
S04 mouth A GN group: a 3-point curve whose middle point follows a bone (Object Info), Curve to Mesh with a profile, and a gradient attribute from curve parameter → material — GN
S07 Stepped animation as data: keys written straight into F-curves with interpolation = 'CONSTANT'; holds chosen per action (2s for gags, 3–4s for settles); expression index keyed the same way a beat sheet (dialogue timing) bridge
S09 EEVEE, Standard view transform; B/W via a desaturating grade (comp Hue/Sat at 0); a spotlight “comedy club” key with a hard shadow — EEVEE
S11 VSE as data: sequence_editor.strips.new_scene or new_movie per shot, with cut points from the beat sheet; title card as a text strip — VSE API (new_scene, new_movie, new_effect present in 5.2.2)

Build order

  1. Write the beat sheet and assertions (section 5).
  2. Generate the atlas, then the head with UV Warp and the expression driver.
  3. Body armature with bendy-bone arms, then the GN mouth.
  4. Keys from the beat sheet, stepped.
  5. Three camera angles = three scenes or three cameras with markers.
  6. Render the shots, assemble them in the VSE, render the edit.

4. What works today vs. what needs newblender

Section titled “4. What works today vs. what needs newblender”
  • Works today with the live bridge:
    • The UVWarpModifier type, image pixel writes, drivers.
    • F-curves with CONSTANT interpolation, bendy-bone properties.
    • GN curve groups, EEVEE.
    • VSE strip creation (StripsTopLevel.new_scene / new_movie / new_effect, confirmed on 5.2.2 headless).
    • Declared modifiers plus render plus screenshot, as proven by demo 001.
  • Painful today:
    • Rigify generation is an operator that expects an active meta-rig in Pose mode. It is mode-gated (L-001) and context-gated (L-004). Fitting a meta-rig to a mesh is placement by eye (L-009). Rigify is not enabled in a factory boot (probed), which is another hidden-default trap (L-011).
    • Hand-keyed stepped animation is fine as data, but there is no “result” back. Verifying holds means re-reading F-curves (L-003).
    • Scripted edits leave no undo (L-008). When the owner says “make that gag hold one frame longer”, the agent re-writes the keys instead of stepping back.
  • Not feasible today:
    • Comic performance: acting choices, the take, the timing of a punchline.
    • Lip-sync to Italian dialogue at production quality without a phoneme-to-tile mapping. Feasible with an external aligner, but not built.
    • Puppet and character design in the show’s style.

5. Verification plan (RFC 0001 R6, six layers)

Section titled “5. Verification plan (RFC 0001 R6, six layers)”

For the first slice: 10 s at 24 fps, 3 shots.

Layer Check Threshold
1 Validity Head and body meshes: manifold, normals out, UVs present (demo 001 checks) 0 failures
1 Validity Atlas: correct size and no transparent holes in tile regions 5 tiles; α = 1 over ≥ 99% of tile pixels
2 Spec Every animated F-curve uses stepped keys 100% of keyframes CONSTANT
2 Spec Hold lengths (frame gaps between consecutive keys on body channels) every gap ∈ {2, 3, 4} frames, except deliberate holds ≥ 6
2 Spec Expression swaps: the UV Warp offset only ever lands on a tile boundary offset ∈ {0, 0.2, 0.4, 0.6, 0.8} ± 1e-6 at every frame
2 Spec Rubber-hose arm: bendy bones with no kinks bbone_segments ≥ 6; max angle between adjacent segments ≤ 35°
2 Spec B/W max per-pixel chroma (HSV saturation) ≤ 0.02 on all final frames
2 Spec Edit exactly 3 strips; cuts at beat-sheet frames ± 0; total 240 frames
3 Reference Stepped motion actually reads as stepped: the frame-to-frame pixel difference is zero on hold frames ≥ 90% of in-hold frame pairs have mean abs diff < 0.5/255 in the character mask
4 Downstream Animator-friendliness: one “expression” integer property drives the face; no keys sit directly on the modifier; the rig is posable with no errors at extreme arm poses 1 control; 0 direct keys; bendy arm length change ≤ 5% at ±150°
5 Appearance (warning) Vision model: “reads as paper cut-out / stop-motion puppet, black and white comedy” warn if below 3 of 5
6 Taste (owner) Is the beat funny? Does the timing land? owner

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • Comedy. Timing a punchline, choosing when to hold. The tooling can enforce “on twos”, but not “funny”.
  • Puppet and tile design. Five expressions that read instantly are a drawing skill. Procedural ellipses are placeholders.
  • Acting and lip-sync choices for dialogue-driven scenes.
  • Direction across 6 episodes: tone, the colour-episode decision, edit rhythm.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, ~1 day): on screen in the owner’s Blender:
    1. A paper-textured box head appears on a stick-figure body with long noodle arms.
    2. The face changes instantly as the agent sets expression = 0…4 (neutral, smug, shocked, sad, laughing) through the UV Warp tiles.
    3. Scrubbing the timeline, the puppet moves in stop-motion jerks, on twos for the gag and threes for the settle. The arm bends in a smooth rubber-hose curve.
    4. Rendered view goes black-and-white under a single spotlight.
    5. The VSE shows 3 strips cut on the beats, and a 10-second MP4 plays back with the check report.
  • Full reproduction (a 30 s scene):
    • Agent: about 30–60 hours, including a Rigify-based character and a lip-sync mapping.
    • Compute: minutes per shot of EEVEE.
    • Human: 10–20 hours of direction, design and acting passes.
    • The series: 100 minutes at the original’s 40 s/week peak with 20 people. Agents could absorb rig setup, stepped-key cleanup, render wrangling and conform, not authorship.

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • Whether pose.rigify_generate runs through the bridge without GUI context is untested.
  • The source is a single user story. Episode-level breakdowns, the exact UV Warp setups and the GN mouth graph are not public, so the agent’s versions are reconstructions.
  • Stepped renders mask in-between errors that would show on spline playback. That is intended, but it can hide rig bugs.
  • “Prime Video original” and the 3 June 2025 date come from the index (Wikipedia via search), not from the fetched user story.

9. Sources

Section titled “9. Sources”
  • blender.org: Creating “Il Baracchino” (fetched 2026-10-07: ~20 people, 6 eps / 100 min, 3.6 LTS, EEVEE, VSE, UV Warp, Leonardo’s 5 tiles, Rigify + bendy bones, GN mouth, stepped twos/threes/fours, 40 s/week, Kitsu/Flamenco/SVN, GP previz, B/W with one colour episode)
  • Wikipedia: Il Baracchino (via the index, verified via search)
  • Local probes (headless 5.2.2, factory startup): UVWarpModifier present; VSE StripsTopLevel has new_scene, new_movie, new_effect; Rigify not enabled by default
  • docs/discovery/ledger.md L-001, L-003, L-004, L-008, L-009, L-011; RFC 0001 R6; bridge: tools/live/bl.py, snap.py, demo 001 report