Imported from Quriosity-agent/qcut (qcut/.agents/skills/qcut-toolkit/jianying-video-effect-reference/SKILL.md). Install upstream with npx skills add Quriosity-agent/qcut --skill jianying-video-effect-reference. Copyright stays with the author.

Jianying Video Effect (画面特效) Reference

Treat a Jianying 特效 as a four-layer relationship:

catalog card (effects2 / face-prop panel, http_cache JSON)
  -> materials.video_effects[] material          (draft)
  -> own segment on a dedicated type:"effect" track
  -> Cache/effect/<effect_id>/<md5>/ AmazingFeature package
  -> single-input, seek-mode, event-parameterized renderer

This complements jianying-reference (generic package harvesting + stepped-frame capture — reuse its mtime-marker protocol and capture traps verbatim) and parallels jianying-transition-reference (transitions are the dual-input case; 特效 are single-input).

Layer 1 — Catalog (ressdk_db http_cache)

The 特效 panel is TWO API panels: effects2 = 画面特效 (effect_type: 7) and face-prop = 人物特效 (effect_type: 8). All catalog state lives in the http_cache table (raw response JSON keyed by hashed request URL); every structured table (effect, category_effect, panel_*, loki_*) is empty in current builds — do not query them.

~/Movies/JianyingPro/User Data/Cache/ressdk_db/rp_master.db   # ressdk_db_info: hash_path -> (did, uid)
~/Movies/JianyingPro/User Data/Cache/ressdk_db/<hash>/rp.db   # one per (device, uid); uid=0 = logged out

http_cache(id, url, response_body, version, timestamp) — two URL families:

• /artist/v1/panel/get_panel_info_<32HEX>__jianyingpro_{0|beta}_… — panel category list + page 1 of the default category embedded in data.category_resources. The 32-hex token is a hash of the full request params, so the panel name is NOT in the URL; find the effects2 panel by matching data.categories[].category_id against item category_ids.
• /artist/v1/effect/get_resources_by_category_id_<32HEX>_<panel>_… — one 50-item page. Panel name IS readable in this URL suffix (effects2, face-prop, filter, transitions, …). data.next_offset advances 50→100→…, data.request_id stays constant per browsing session, has_more:false marks the last page.

effects2 categories observed (2026-08, 28 tabs, category_id 中文名 key): 39654 热门 rm · 7728 基础 basis · 7730 动感 cool · 7729 氛围 dream · 43957 最新 new · 38510 潮酷 chaoku · 7735 边框 zoomout · 5914834 多屏 · 5914631 有声 · 39547 光 light · 21924 爱心 heart · 5914352 音频 · 5913855 创意AI · 5913856 运镜 · 7734 自然 reality · 39241 金粉 jinfen · 7731 复古 retro · 39246 电影 dianying · 15502 Bling kira · 39539 扭曲 niuqu · 27966 综艺 zongyi · 7733 分屏 split · 5914473 宠物 · 5913770 投影 · 5913775 纹理 texture · 7732 漫画 comic · 37381 暗黑 halloween · 39264 DV dv. Beware: face-prop has its OWN 热门 (38389) — a draft category_id tells you which panel the card came from.

Per-item fields that matter (data.effect_item_list[].common_attr unless noted):

Top slider keys by frequency: effects_adjust_speed (545/680), intensity, luminance, blur, background_animation, filter, size, color, range, distortion, horizontal_shift, vertical_shift, horizontal_chromatic, sharpen, … — normalized 0–1 in catalog and draft (~99.5%; a handful declare real ranges, e.g. luminance max 2.3).

Coverage caveats: only categories the user actually browsed have cached pages (one row per 50-item scroll page); packages download lazily on FIRST APPLY, so a machine can know ~1000 catalog effects yet hold only ~10 on disk (2026-08-16 measurement: 1024 rows — 707 effects2 + 317 face-prop — of which 618 are blit-only). To widen the CATALOG, browse tabs in the app; but packages themselves no longer need in-app clicking — item_urls[0] is directly fetchable and md5 verifies the zip (see the batch pipeline section below).

Layer 2 — Draft (materials.video_effects + effect tracks)

Current draft_info.json is encrypted; evidence comes from plaintext */.backup/*.load.bak files (rare — scan with head -c5 | grep '{').

An applied 特效 = one materials.video_effects[] entry + one segment on a dedicated type:"effect" track. The segment's material_id is the material UUID and its extra_material_refs is []. Video segments NEVER reference video_effects (contrast materials.effects — filters/adjusts — which can be clip-attached via extra_material_refs). Material schema:

{
  "id": "<UUID>",                    // what segment.material_id points to
  "type": "face_effect",             // observed for face-prop; 画面特效 presumably "video_effect" (unverified)
  "name": "发光分身",
  "effect_id": "13661053",           // Cache/effect/<effect_id>/ dir (legacy short id — NOT the catalog effect_id)
  "resource_id": "7233250530292666939", // = catalog third_resource_id_str; ≠ effect_id (filters have them equal)
  "path": "…/Cache/effect/13661053/dbff6732319cf9488da4816c188d9f1a", // container path symlinks to ~/Movies
  "category_id": "38389", "category_name": "热门",   // which panel tab it came from
  "adjust_params": [{"name": "effects_adjust_intensity", "default_value": 0.8, "value": 0.8}],
  "apply_target_type": 2,            // only value observed; effects[] filters use 0.
                                     // Community convention 0=clip 2=global — NOT verified locally.
  "apply_time_range": null, "time_range": null,      // timing lives on the segment
  "value": 1.0,                      // overall strength
  "algorithm_artifact_path": "##_draftpath_placeholder_<GUID>_##/video_effect/multi_faces_algorithm/<UUID>",
  "disable_effect_faces": [], "common_keyframes": []
}

Effect segment specifics: universal segment schema with source_timerange: null, clip: null (generated content, no media/transform), target_timerange in microseconds, default duration: 3000000 (3 s, matching the catalog's effect_duration), start = playhead at apply time.

Render-order bands (render_index_track_mode_on: true): main video 0, PIP 2, filter tracks 10000+, effect tracks 11000+ (one +1 per effect segment in creation order), sticker 14000+. So 特效 composite above filters and below stickers. track_render_index is just the track's array index. adjust_params[].default_value mirrors the package extra.json defaults exactly; value is the user's slider state.

Layer 3 — Package (Cache/effect AmazingFeature bundles)

Cache/effect/<effect_id>/<md5>/ — shared by transitions/masks/beauty too (filters live in Cache/artistEffect/ instead). Two md5 dirs can coexist under one effect_id (old + updated package versions — a draft's path may pin the older md5 while the catalog already lists the newer one); a sibling <md5>_modity_time.txt (present on the most recently downloaded version) lists the extracted file manifest.

Every 特效 package is an Amazing-engine bundle (NOT the .lsproj node graph — in this cache that format belongs exclusively to text animations):

config.json           # effect.Link[] {path, type:"AmazingFeature", zorder≈8011}; bALG_BACH_CONFIG
extra.json            # setting.effect_adjust_params — same schema as catalog sdk_extra
algorithmConfig.json  # optional CV graph: nodes[] {type: blit|face|matting, config}, links[]
AmazingFeature/
  main.scene          # binary %SerializedFormat% scene
  sticker.config      # engine systemList + dev/min_version gate
  material/ rt/ mesh/ # entity-material graph, render targets
  xshader/*.frag      # PLAINTEXT GLSL, and/or compiled plaintext pairs under
  Library/ShaderData/<hash>/shaderGLES|shaderMetal/
  lua/SeekModeScript.lua, ImageBusinessSlider.lua, …

Renderer families observed (classify before porting):

Render contract (the part QCut parity work must honor):

• Single input texture — shaders sample one inputImageTexture; Lua reads Amaz.BuiltinObject:getInputTextureWidth/Height(); Lumi scripts get InputTex/OutputTex/PingPongTex. Extra samplers are internal RTs or CV masks (matting masks arrive y-flipped, value in .r), never a second clip.
• Seek-mode time — families (a)(b)(d)(e) carry a SeekModeScript.lua; progress = (curTime - startTime) / (endTime - startTime) with endTime = 3.0 in most packages (a few use 10.0/0). The host-write mechanism varies: some scripts declare an explicit --@input float curTime slider, others are driven through an autoplay/playTime field. Family (c) Lumi/AE packages have NO SeekModeScript — lua/LumiFamily/LumiManager.lua accumulates deltaTime in onUpdate (default endTime 6.0) and pushes start/end/curTime down the layer tree, so they are wall-clock-stepped, not closed-form scrub-safe. Verify which contract a package uses before assuming QCut-style f(progress) parity.
• Params as events — the host dispatches onEvent(key = "effects_adjust_*", value ∈ [0,1]) with the draft's normalized value. An auto-generated ImageBusinessSlider.lua (header write by editor EffectSDK) remaps to real uniform ranges declared data-side in AmazingFeature/ImageBusinessSlider.json (per-key → {entity, material, uniform, minValue, maxValue}; e.g. 发光分身 intensity→u_GlowIntensity@Blend [0,1], size→u_Scale.y@displace0 [0.05,0.5]). So draft 0–1 values are meaningless without the package's remap table — harvest it before calibrating.
• CV pipeline — requirements + model_names (catalog) declare it; algorithmConfig.json is the executable graph (blit downsample size, face detect ability flags). Effects needing matting/face cannot be ported as pure shaders; QCut needs its own segmentation/landmark source or a documented known difference.

Rendering a package through the local runtime (verified)

QCut can render these packages by driving the Jianying runtime installed on the machine — the same libraries the Transition Lab uses. Decoded by reading each created object's vtable pointer back to its symbol:

SwingSegmentType: 0 = FeatureSegment (特效), 2 Sticker, 3 Text, 4 Template, 5 Emoji, 6 Custom, 7 Video, 8 Transition, 9 StickerBrush, 10 Script. 1/11/12 are unmapped and crash — the factory bounds-checks <= 0xa.

Render contract (proven across every family above, including Lumi/AE):

1. bef_swing_manager_create_with_gpdevice inside a GL/Metal context, then set_parameter_bool(manager, "EnableSwingSimplify", true) — required, or nothing renders at all.
2. Video segment = type 7; effect = type 0 created WITH the package path (that is what loads main.scene).
3. bef_swing_segment_video_add_feature(video, feature) — a 特效 is a feature ON the video segment, not a standalone segment. Adding it to the manager as well trips _preProcessWithoutTracks: invalid segments, two -1 layer found and voids both segments.
4. Per frame: video_set_device_texture(video, &input) then manager_seek_frame_device_texture(manager, timestampMicroseconds, &input, &output).
5. Sliders arrive as effects_adjust_* key/value pairs (normalized 0–1, straight from the draft).

The trap that wastes hours: most effects are IDENTITY at most timestamps — 抖动 differs only near 0.2 / 0.6 / 1.2 s. A single-timestamp test reads as "the effect never rendered". Sweep time, or validate against an always-on overlay (胶片框 / 怀旧边框 II hold a constant ~17–19 mean channel difference).

The runtime needs the FULL 23-library closure (~/Library/Application Support/qcut/PrivateRuntimes/JianyingTransition/current); the 5-library .local/jianying-runtime root segfaults during manager init.

Implementation: research/jianying-runtime-probe/effect-probe.mm + electron/jianying-effect/ (PR #414).

Batch reference pipeline (参照生产线)

scripts/jianying-effect-reference-batch.cjs mass-produces ground-truth reference clips for every blit-only catalog effect — the scale path chosen 2026-08-16 (the lab itself cannot scale to end users: macOS ∩ JianYing installed ∩ package cached; references drive QCut-native reimplementation instead, like the filter LUT fitting and sound-effects lab precedents).

bun run build   # dist/electron must be fresh — the script requires it
node scripts/jianying-effect-reference-batch.cjs              # full run
node scripts/jianying-effect-reference-batch.cjs --limit 5    # smoke
node scripts/jianying-effect-reference-batch.cjs --panel effects2
node scripts/jianying-effect-reference-batch.cjs --only <effectId,...>

Mechanics worth knowing before touching it:

• node, never bun — the catalog reader statically needs node:sqlite. The dist/electron/jianying-effect/* modules have zero electron imports, so plain node can require() them (getFFmpegPath resolves the staged dev ffmpeg).
• Acquisition: reuses packages already in JianYing's caches (md5-indexed), else fetches item_urls[0] (UA-spoofed, 500 ms throttle), verifies md5(zip) == common_attr.md5 == dir name, unzips into .local/jianying-effect-references/_packages/<effectId>/<md5>/. JianYing's own cache is never written to — definition.packagePath accepts any directory.
• Render: 6 s / 1280x720 / 30 fps via renderJianyingEffectClip, default slider values from the package extra.json (fallback catalog sdk_extra), effect window spanning the whole clip. ~5-6 s per effect cached, 8-13 s with download.
• Reference clip: real footage (skate) + SMPTE bars + grayscale ramp strips — fit on real colors only (LUT-fitting lesson).
• Manifest: manifest.jsonl append-only, one line per effect (md5/seconds/frames/ssim/adjust params/error); reruns skip ok:true rows, so interrupted runs just resume.
• SSIM ≠ identity check: sparse effects legitimately score ~0.99 (星火); only >0.997 gets flaggedIdentity, and even that needs a frame sweep before concluding (same trap as the single-timestamp one above).
• The Lumi "failures" were a truncation artifact, not a runtime limit. A 2026-08-17 run reported 517/618 with all 101 failures in the Lumi family (~90 the JS variant: root LumiManager.js + config.json.js_path + embedded ThreeJS; rest Lua LumiFamily/), and the family's known wall-clock/onUpdate contract made "needs native stepping" look obvious. Running the bridge directly disproved it: exit 0, a complete raw output, the [effect] frames: receipt present in stdout, and 60/60 frames differing from the input. The receipt sat 10,268 bytes from EOF while render.ts kept only the last 8,192 — Lumi's JS-engine teardown ([AE_JSRUNTIME_TAG]'Scene: 开始清理场景资源' plus dispose/destroy stacks, ~90KB) scrolled it out. Packages without a JS engine barely log at teardown, which is why the false failures landed exactly on one family. Fixed by latching the line as it streams (retainPattern); all 101 then passed, taking the library to 618/618.
• Diagnostic rule this cost us: when the bridge "finishes without reporting", read the FULL raw stdout before theorising about the render contract. Tail-window truncation is a silent source of false failures. (A failed pass still muxes an mp4 — the third ffmpeg step runs before the receipt is parsed — so the batch script deletes its output on failure.)
• The lab UI consumes manifest.jsonl as a verification ledger: effects with a failing verdict are locked as 「本机渲染验证未通过」 instead of pretending to work (electron/jianying-effect/catalog.ts).
• Outputs live in .local/jianying-effect-references/ (gitignored). Same red line as everything else here: packages and rendered references never enter git, public storage, or the product. Team sharing goes through the sound-effects-lab private-bucket + allowlist pattern.

Companion doc: docs/task/jianying-effect-reference-line/README.md.

Harvest protocol

1. Map card → package with the mtime-marker loop from jianying-reference (apply ONE card, find the new Cache/effect dir; already-cached cards leave zero disk trace).
2. Read extra.json + ImageBusinessSlider.json first (param schema + remap ranges), then classify the family (table above) before reading shaders.
3. Compiled shaders under Library/ShaderData/ are plaintext even when .ausl sources are encrypted — same trick as text animations.
4. Capture stepped reference frames per jianying-reference's protocol (loop effects need the play-pause capture variant).
5. All of jianying-reference's Capture Traps apply unchanged.

QCut landing zone (2026-08 state)

• Panel EXISTS: apps/web/src/components/editor/media-panel/views/effects.tsx (tab effects, label 特效, flag VIDEO_EFFECTS in config/features.ts), ~166 published presets in apps/web/src/lib/effects/effect-catalog.ts (+13 per-kind catalog files).
• Render model: EffectRenderProgram with 9 stage kinds in packages/editor-core/src/types/effect-render.ts; stage types are DUPLICATED in electron/ffmpeg/effect-render-types.ts — mirror both when adding a kind. Preview seam: preview-panel/use-effects-rendering.ts + preview-element-renderer.tsx. Export seams: canvas (lib/export/export-engine-renderer.ts) and CLI/FFmpeg (lib/export-cli/sources/effect-*-sources.ts).
• Timeline: active mechanism is per-clip element.effects: EffectInstance[] (synced with stores/ai/effects-store.ts); TrackType "effect" and EffectElement types exist in editor-core but nothing constructs them — Jianying's independent draggable effect clips are a gap.
• CapCut draft export does NOT emit materials.video_effects (unsupported-features.ts raises an error for element effects); import maps the bucket to segment kind effect with capability opaque. Both are open parity work.
• Prior research to reuse: docs/task/effects-pack/ (JIANYING-LIST.md, MAPPING.md, PRIMITIVES.md, CHECKLIST.md) — panel screenshots, 44 mapped gaps, stage capability boundaries, add-one-effect recipe.

Scope notes

• Read-only analysis of locally cached files for interop/parity. Do not redistribute Jianying assets or ship harvested content — and that covers the REPO, not just the product: no decompiled shaders, prefab dumps, strings output, or extracted media in commits. Transcribe behavior into your own equations/prose; raw files stay in the session scratch dir.
• Numbers above (category ids, counts, band values) were measured on the 2026-08 CN build (draft version 360000, app 5.9.x); re-verify ids after app updates — catalog hashes embed the app version, so stale rows linger beside fresh ones in http_cache.