Iteratively improve Blender music videos with targeted still probes, contact sheets, segment renders, video-only outputs, and final audio muxing. Use when refining generated MP4 visuals, debugging visual artifacts, checking timing against music, or when the user asks to inspect/fix a specific moment in a Blender video. For low-poly rigged character limb spike/fin artifacts, see references/rigged-game-asset-pose-iteration.md.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Iteratively improve Blender music videos with targeted still probes, contact sheets, segment renders, video-only outputs, and final audio muxing. Use when refining generated MP4 visuals, debugging visual artifacts, checking timing against music, or when the user asks to inspect/fix a specific moment in a Blender video. For low-poly rigged character limb spike/fin artifacts, see references/rigged-game-asset-pose-iteration.md.
Use this skill for the practical loop of making generated Blender music videos better: measure the song, render targeted stills, build contact sheets, inspect visually, patch one problem at a time, and only then render/mux an MP4.
The rule: do not claim a visual fix from code alone. For every visual correction, produce real frames or a real MP4 segment and inspect them.
When the video features a person/humanoid character, load blender-character-video first; character work needs a face/body/pose approval loop before ordinary scene iteration.
Detailed commands live in references/workflow-recipes.md. For person/humanoid-centered videos, use references/person-character-video-lessons.md: run a character-first approval loop around face/body/pose/outfit/accessories, promote the exact approved rig/texture path into the final generator, and motion-probe before full render. When a late-stage WAV changes, the user asks whether audio clips, asks for a small loudness bump, or notices a blink/eye state lingering, use references/audio-loudness-and-blink-polish.md: measure the current WAV, make a named louder master with safe peak headroom, mux it last, verify the final AAC, and treat half-open eyes as quick transition frames unless explicitly desired. For imported low-poly/game character rigs, use references/rigged-game-asset-pose-iteration.md before posing limbs or claiming a character fix. For rail/surface hand-tap fixes on low-poly characters, use references/low-poly-hand-tap-iteration.md: translate the IK target vertically and verify down/neutral/up, rather than rotating the palm toward camera. If the user means literally only fingers/fingertips tapping, use references/finger-only-low-poly-taps.md: keep the arm/IK pose fixed and animate only distal fingertip vertices. If the user adds a new character texture/eye set while refining fingertip contact, use references/low-poly-texture-eye-and-finger-contact.md: update the main atlas and separate blink eye overlays, then verify fingertip press direction with close crops. If the user wants to use an image generator to repaint a UV atlas, use references/low-poly-uv-atlas-repaint-prompts.md: inspect the real FBX UVs, generate a labeled overlay, and prompt with exact atlas-coordinate constraints. If the user rejects bones/masks and wants exact visual cutouts for ChatGPT/image editing, switch to references/rectangular-cut-chatgpt-uv-workflow.md: generate code-cut PNG tiles, a contact sheet, manifest, prompt, and paste-back script before further explanation. If rectangular guesses are too coarse or the user asks whether UV parts are really rectangles / wants a saner export, use references/uv-island-chatgpt-cut-pack.md: generate exact FBX UV-island crops with overlay/alpha files, numbered overview, SVG, manifest, and paste-back helper. If an image-editing API/key is unavailable or the generator is too unconstrained, use references/local-uv-atlas-repaint-fallback.md: make a deterministic, coordinate-bounded local texture pass, generate matching eye overlays, inspect a texture sheet, and render a still. When mixing imported-character texture relinks with online sky/moon/star assets, use references/gyre-nasa-sky-assets-and-texture-scope.md: keep NASA/sky images in a separate folder, use full-disk moon/starfield assets when hand-drawn moons/stars fail, and scope character atlas relinking so it never overwrites environment materials. If moonlit “clouds” read as chunky puffs or ellipses and the user asks for mist, use references/gyre-mist-vs-cloud-puffs.md: replace the visual language with broad, feathered, low-opacity alpha mist cards and inspect the same approval still for card boundaries. When generating candidate atlases or eye/blink overlays, use references/uv-atlas-generation-asset-safety.md: preserve current assets for comparison, generate into attempts folders, and install only after explicit approval. When using rectangular texture cuts plus masks, use references/crop-mask-candidate-texture-probes.md: call it crop + paste-mask, keep it candidate-only, temporarily install only for one Blender still, then clean up and verify no new_* assets remain installed. If UV islands/crops are technically correct but the user questions whether they are semantically accurate enough for hair/clothing/skin regions, switch to references/texture-paint-semantic-atlas-workflow.md: create a paint-ready .blend, apply the current atlas to the real imported asset, aim the camera at the mesh bounds, render a preview, and let Texture Paint map brush strokes back to exact atlas pixels. If a costume/skin cutout renders as W-shapes, doubled seams, stacked diamonds, or other UV-island artifacts after repeated atlas edits, use references/model-space-character-texture-fixes.md: restore the atlas area to clean base texture, drive the visible mask from object/model-space material coordinates, rerender the same still, and inspect before claiming the fix. If a low-poly character needs a real goth/witch palette pass — different outfit, paler skin, different hair and eyes — or atlas-painted chest details keep splitting into W/diamond artifacts, use references/low-poly-character-goth-palette-and-neckline.md: update the whole character read, include eye overlays, and use object/material-space masks for single clean cutouts when UV islands fight the shape. If the user asks for a forward rail/broom lean with head tilted down, use references/low-poly-forward-lean-rail-pose.md: pose the spine chain, counter-rotate the head enough to keep eyes visible, keep hands anchored, and make hats/accessories follow the posed head. If the user then asks for small character-art polish such as a larger bust, a hat moved up/down, or top hair showing through the hat, use references/low-poly-character-accessory-proportion-polish.md: adjust the real mesh/accessory placement, tuck only offending top/front hair vertices, rerender the same still, and inspect face/eyes/hair/clipping before claiming the fix. If the remaining weak point is shirt/corset/costume lookdev and the user asks for several creative passes before another video render, use references/ps1-character-shirt-lookdev-options.md: render a labeled still-only option sheet from the final camera, recommend top picks, and wait for explicit selection before full MP4 rerender. If an imported anime/VRoid-style character has baked-in underwear/test clothing and the outfit pass starts turning into tall panels, use references/fitted-clothes-on-imported-anime-character.md: clean/replace the base body material first, then add short fitted wardrobe shells, and inspect before posing.
When a single approved character still needs to become the full music video, use references/approved-character-still-to-final-video.md so the main generator actually uses the approved rig/path before final render. If the requested final is mostly a repeating blink/tap/twinkle/mist motion cycle over a long song rather than unique full-song camera progression, use references/approved-still-to-looped-full-song-video.md: render and inspect a short loop, encode it with ffmpeg -stream_loop -1 to the source duration, mux audio separately, and disclose/document that the final is a looped visual cycle.
When to Use
Use when:
The user points to a timestamp or screenshot and asks why a visual looks wrong.
A music-video render needs iterative art direction: sky, stars, road lines, parallax, camera motion, fog, sun, shader planes, etc.
You need to verify whether audio-reactive visuals actually read in the final video.
You need to splice a corrected visual segment into an existing render.
The user asks for video first and audio merged afterward.
If a low-poly/game-character import needs visible pose polish, arm/hand fixes, or first-frame art-direction iteration.
A user rejects bone-derived masks / heuristic masks and wants to manually cut or paint UV atlas regions for ChatGPT/image-edit replacement; use references/manual-uv-cut-paint-workflow.md.
Do not use for pure Strudel/music parsing; use strudel-test for that.
Use additional named outputs for experiments, e.g. final_linefix.mp4, final_cloudfix.mp4, cloudfix_segment_24fps.mp4.
Core Loop
Measure and anchor the timeline. Use ffprobe and feature JSON counts. Convert seconds to render frames before selecting probes.
Render targeted stills. Probe the exact moments the user criticized and nearby frames, not only pretty overview frames.
Create a contact sheet. Combine stills into one image and inspect it with vision. Ask a specific visual question.
Diagnose from evidence. Match visible artifacts to likely causes before patching.
Patch one thing. Keep camera timing, sky geometry, audio reactivity, rig/pose work, and mux/export changes separate. When the user is approving a single character still, overwrite the same named still until approved instead of creating probe-folder sprawl.
Respect approval gates. If the user is still giving still-frame art direction, says “keep going” in the context of a still, asks for options/passes, or says “don’t do the whole video until I tell you,” do not render frame sequences, final video-only MP4s, mux audio, or write final-delivery README updates. Stay in single-still/contact-sheet mode until they explicitly choose/approve a direction and ask to render the video. Ambiguous enthusiasm like “we good to go now?” is not approval if the user also asked to inspect options.
Render a motion segment if needed. Stills prove composition, not beat sync or travel.
Make a silent video first. Splice or render *_video_only.mp4 only after the still/segment gate is approved.
Mux audio last. Add source.wav via ffmpeg in a final pass. If the user replaced the WAV or asks for louder audio, first measure clipping/loudness and create a named safe-gain master (for example source_louder.wav) rather than silently changing source.wav.
Verify and document. Run ffprobe, inspect final contact sheet, update renders/<song>/README.md after final render/mux work is actually requested. For louder finals, verify both the louder WAV and final AAC peaks/true peaks; for blink changes, inspect an eye-focused sheet/crop.
Diagnosis Patterns
Emission changes but motion appears stopped: check camera/travel curves for clamps. Convert clamp progress to wall-clock time (progress * duration). A curve clamped at 0.80 stops at 168.8s in a 211s song.
Sky/moon image assets showing character textures: a broad character texture fix may be rewriting every loaded image/material after the environment was created. Scope atlas relinks to imported character materials only and skip NASA/sky/moon/starfield/cloud materials. Use full-disk moon mosaics for hero backdrops instead of equirectangular moon maps on spheres when the still needs an iconic readable moon. See references/gyre-nasa-sky-assets-and-texture-scope.md.
Clouds become ellipses during spin: they are likely flat cards or heavily flattened ellipsoids. Use more small rounded puffs/star knots and reduce individual self-rotation; rotate positions around the sky dome instead. If the user says the clouds suck or asks for mist, stop iterating cloud puffs and switch to broad, feathered, low-opacity mist ribbons/cards; inspect for rectangular card edges and beam-like opacity. See references/gyre-mist-vs-cloud-puffs.md.
Sky/bands look tiny: inspect camera projection; far geometry may project into the center only. Make skybox panels camera-relative or dramatically larger/nearer.
Stills look fine but motion feels wrong: render a short silent segment around the problem.
Audio-reactive code exists but is not visible: inspect actual MP4 frames; material changes can be swamped by exposure, distance, or camera motion.
Late replacement WAV / louder request: re-measure the current source.wav for peak, true peak, and loudness before applying gain. Use a conservative named louder master with headroom, mux it last, then verify the encoded final AAC; see references/audio-loudness-and-blink-polish.md.
Blink half-state appears stuck: a half-open texture can read as a held expression if it appears on both entry and exit. For normal blinks, make half-open a one-frame/quick transition (open -> half -> closed -> open) and verify with a close eye sheet/crop. If the user wants half-open/closed eyes sometimes, add separate intentional acting holds on a slower cycle rather than reintroducing a return-half blink linger; see references/audio-loudness-and-blink-polish.md.
Low-poly blink/tap motion probes
When an approved low-poly/game-character still needs a short motion probe, keep the exact approved clean-rig path and animate controls rather than mesh vertices. For hand tapping, move the relevant IK target vertically in world Z; do not rotate the hand target if that makes the palm face camera, and do not lift only edge vertices if that creates spike/fin artifacts. For texture-based blinks, swap eye textures per rendered frame and verify with a contact sheet. See references/low-poly-rig-blink-tap-loop.md.
Segment and Mux Discipline
When only part of a long video needs correction:
Render the corrected range as a silent segment.
Keep full-song progress mapping with --timeline-start-frame 1 --timeline-end-frame <full_end_frame>.
Splice the corrected segment into the previous video as final_video_only.mp4.
Mux source.wav into the final MP4 with -map 0:v:0 -map 1:a:0 -c:v copy -c:a aac -shortest.
Treating a still as proof of motion. Stills prove composition, not beat sync or travel.
Forgetting timeline mapping during segment renders. A segment can accidentally evaluate as progress 0.
Muxing audio before visual debug is finished. Keep visual iterations silent; add audio at the final step.
Changing too much at once. It becomes impossible to know what fixed the problem.
Letting generator comments lie. If the code says a thing is 3D but it renders as cards/lozenges, update code and comments.
Ignoring user screenshots. Reproduce that timestamp and inspect adjacent frames.
Stopping after a remote command times out. Check logs and frame counts; the render may still be running or may have usable frames.
Imported game rigs need control rigs, not hacks. Preserve the chosen asset; avoid procedural limb/costume overlays, manual vertex remapping, and unverified Euler FK. Build/helper IK controls with pole targets, or rebuild a clean armature named to the original vertex groups and bind it to the existing weights, then inspect pose sheets/close stills before claiming arm or hand fixes. See references/clean-vertex-group-retarget.md.
Probe sprawl during single-frame approval. If the user is approving one still, keep overwriting the agreed path (for example renders/<song>/one_frame.png) instead of making new folders/contact sheets/multiframe runs unless explicitly requested.
Final-render eagerness violates lookdev. A positive reaction to a still (“way better, keep going”) means continue the current approval loop unless they explicitly ask for the whole video/final/mux. Do not infer final-render permission from enthusiasm.
Finger-only means finger-only. If the user rejects arm/hand bobbing and asks for only fingers to move, lock the arm/IK target and animate only a small distal fingertip vertex region. Verify with a close crop that includes the forearm and wrist; a full-frame sheet can hide unintended arm motion. See references/finger-only-low-poly-taps.md.
Vertical hand taps are translations, not palm rotations. If a low-poly hand should tap a rail/surface up and down, moving/rotating the palm can make it face the camera while only appearing to fix spikes. Keep the rail-facing hand orientation and animate the IK target’s vertical Z offset; prove it with a down/neutral/up contact sheet. See references/low-poly-hand-tap-iteration.md.
Texture swaps may need eye overlay swaps too. Low-poly characters can use a main body atlas plus separate open/half/closed eye textures. When a user provides a new atlas, look for matching new_bg_eye*.png overlays and update blink helpers to prefer them. See references/low-poly-texture-eye-and-finger-contact.md.
Finger contact direction matters. If the wrist is already touching the frame/rail, do not lift or move the hand. Press only distal fingertip vertices downward toward the surface and verify with a close crop. See references/low-poly-texture-eye-and-finger-contact.md.
Generated atlas repaints need UV proof, not vibes. Before asking an image model to make a low-poly character texture more witchy/gothy/etc., inspect the actual FBX UVs, draw a labeled overlay on the atlas, and prompt with exact coordinate regions plus “preserve atlas layout” constraints. See references/low-poly-uv-atlas-repaint-prompts.md.
If the user wants cuts, generate cuts. Do not keep arguing bones/masks when the user asks for exact image pieces. First produce rectangular code-cut tiles if that is enough; if they ask whether the pieces are really rectangles or want a saner export, produce exact UV-island carrier crops with overlay/alpha files. See references/rectangular-cut-chatgpt-uv-workflow.md and references/uv-island-chatgpt-cut-pack.md.
Texture candidates are not installed assets. Preserve the current atlas and eye overlays for side-by-side comparison. Generate provider/image-model candidates into attempts directories, build a compare sheet, and only copy them into new_*/eye asset paths after explicit user approval. Do not silently substitute local deterministic painting when the user requested image-model generation. See references/uv-atlas-generation-asset-safety.md.
Crop is not paste safety. If using rectangular “cuts,” call the method crop + paste-mask. The crop supplies edit context; the mask controls what actually returns to the atlas. Keep edits candidate-only, prove them with a compare sheet plus one Blender still, then remove temporary installed new_* files and verify the live asset folder is clean. See references/crop-mask-candidate-texture-probes.md.
UV accuracy is not semantic accuracy. Exact UV islands can still mix hair, skin, clothing, hands, and empty atlas space. If the user says the pieces are not accurate enough, stop defending atlas cuts and create a Texture Paint workspace on the real imported character so visible brush strokes map back to exact atlas pixels. See references/texture-paint-semantic-atlas-workflow.md.
Continuous costume shapes may need model-space masks. If an atlas-painted neckline/cutout repeatedly renders as W-shapes, doubled seams, stacked diamonds, or panel splits, the artifact is probably UV-island structure. Restore the atlas area to clean base texture, use object/model-space material coordinates for the visible mask, and rerender/inspect before claiming the visual fix. See references/model-space-character-texture-fixes.md.
Global character read beats local detail. If the user says the character needs a different outfit, paler skin, different hair, and different eyes, treat that as a full palette/outfit pass. Do not spend the next turn defending or polishing a neckline-only change. See references/low-poly-character-goth-palette-and-neckline.md.
Texture relinks can clobber environment assets. If a scene imports a character and also loads online moon/star/sky images, a generic for img in bpy.data.images or for mat in bpy.data.materials character-atlas relink can overwrite the sky assets with the character atlas. Filter by path/name/material role and skip sky_sources, NASA, moon, starfield, cloud, rail, and other environment materials before rerendering. See references/gyre-nasa-sky-assets-and-texture-scope.md.
If the user rejects hand-made moon/stars, replace the source language, not just parameters. Use real public-domain/credited sky assets, process them into PS1-friendly cards/maps, and verify that the actual render shows asset detail rather than a procedural approximation. Full-disk moon mosaics often read better for a hero backdrop than equirectangular moon textures on a sphere.
UV seam artifacts can masquerade as costume design. If a chest cutout repeatedly reads as W/straps/stacked diamonds, atlas painting is the wrong representation for that shape. Restore the cloth atlas and use object/material-space masking or real mesh texture paint.
Separate eye overlays are part of the character design. Recolor new_bg_eye01/02/03.png and small compatibility bg_eye01/02/03.png when changing eye color; otherwise the rendered still can keep the old eye color even if the main atlas was changed.
Rail-lean poses need accessory follow-through. For a forward lean with head dipped down, pitch the torso/spine chain forward, counter-rotate the head enough to keep eyes visible, and update hats/hair/accessories from the posed head so they do not float or hide the face. Keep hands anchored on the rail and verify with the same still. See references/low-poly-forward-lean-rail-pose.md.
Hat height, hair clipping, and body proportions interact. Moving a witch hat up can reveal top/front hair through the brim or crown, and body-proportion edits can read differently from the approval camera. Adjust the real mesh/accessory placement, tuck only the offending top/front hair vertices rearward/downward, preserve side locks/eyes, rerender the same approval still, and inspect before claiming the fix. See references/low-poly-character-accessory-proportion-polish.md.
Low-poly “max bust” and arm symmetry require visual probes. If the user asks how large a bust can get, increase real torso deformation in a rendered probe and stop at the visible topology/pose limit, before it reads blobby, clipped, or pasted-on. If one rail-braced arm looks different, mirror math is not enough: probe that side’s IK hand target, elbow pole, and hand roll, then install the best visible still. See references/low-poly-character-accessory-proportion-polish.md.
Louder does not mean normalized-to-zero. If the user asks whether a replaced track clips and wants it slightly louder, measure the current WAV, choose a gain that preserves about 1 dB peak/true-peak headroom, write a named source_louder.wav, mux it last, and verify the final AAC too. Do not claim clipping safety from the pre-mux WAV alone. See references/audio-loudness-and-blink-polish.md.
Half-eye is usually a transition, not a pose. For texture-swap blinks, open -> half -> closed -> half -> open can read as lingering half-lidded eyes. If the user notices it, remove the return half-hold and render an eye-focused contact sheet before claiming the fix. See references/audio-loudness-and-blink-polish.md.
Final artifact folders should not become archaeology. During iteration, keep probes, option folders, frame sequences, and texture attempts named and self-contained; after the final is approved and the user asks for cleanup, preserve deliverables/source/assets/README and remove obsolete probe/frame/option folders with a before/after directory + size check.
Verification Checklist
If using a user-approved character still, the final generator imports the same approved rig/import path rather than an older workaround.
If delivering a looped visual cycle over a full song, the loop was rendered/inspected first, encoded to the source duration with ffmpeg -stream_loop -1, and clearly documented as a loop rather than unique full-song animation.
Source duration and feature frame count measured.
Probe frames target the criticized timestamps.
Contact sheet created and inspected.
If motion is in question, a short segment MP4 was rendered/inspected.
Fix was applied surgically and previous verified fixes were preserved.
Final video-only MP4 exists when using video-first workflow.
Final audio mux completed from source.wav.
ffprobe verifies duration, resolution, frame count, and audio stream.
README.md documents changes, artifacts, and verification evidence.
