mulmocast 2.4.2 → 2.4.4

package/lib/actions/movie.d.ts

@@ -1,7 +1,7 @@
 import { MulmoStudioContext, MulmoBeat, MulmoTransition, MulmoCanvasDimension, MulmoFillOption, MulmoVideoFilter } from "../types/index.js";
 import { FfmpegContext } from "../utils/ffmpeg_utils.js";
 type VideoId = string | undefined;
-export declare const getVideoPart: (inputIndex: number, isMovie: boolean, duration: number, canvasInfo: MulmoCanvasDimension, fillOption: MulmoFillOption, speed: number, filters?: MulmoVideoFilter[]) => {
+export declare const getVideoPart: (inputIndex: number, isMovie: boolean, duration: number, canvasInfo: MulmoCanvasDimension, fillOption: MulmoFillOption, speed: number, filters?: MulmoVideoFilter[], frameCount?: number) => {
     videoId: string;
     videoPart: string;
 };

package/lib/actions/movie.js

@@ -8,7 +8,8 @@ import { MulmoStudioContextMethods } from "../methods/mulmo_studio_context.js";
 import { convertVideoFilterToFFmpeg } from "../utils/video_filter.js";
 // const isMac = process.platform === "darwin";
 const videoCodec = "libx264"; // "h264_videotoolbox" (macOS only) is too noisy
-export const getVideoPart = (inputIndex, isMovie, duration, canvasInfo, fillOption, speed, filters) => {
+const VIDEO_FPS = 30;
+export const getVideoPart = (inputIndex, isMovie, duration, canvasInfo, fillOption, speed, filters, frameCount) => {
     const videoId = `v${inputIndex}`;
     const videoFilters = [];
     // Handle different media types
@@ -21,8 +22,20 @@ export const getVideoPart = (inputIndex, isMovie, duration, canvasInfo, fillOpti
     else {
         videoFilters.push("loop=loop=-1:size=1:start=0");
     }
-    // Common filters for all media types
-    videoFilters.push(`trim=duration=${originalDuration}`, "fps=30");
+    // Normalize framerate first so trim=end_frame counts frames at VIDEO_FPS,
+    // regardless of the input's native framerate.
+    videoFilters.push(`fps=${VIDEO_FPS}`);
+    // Use frame-exact trimming when frameCount is provided to prevent cumulative drift
+    // between video and audio tracks. trim=duration=X rounds up to next frame boundary,
+    // causing ~0.03s extra per beat that accumulates over many beats.
+    if (frameCount !== undefined && frameCount > 0) {
+        // Account for speed: setpts compresses timestamps, so we need more input frames
+        const inputFrameCount = Math.max(1, Math.round(frameCount * speed));
+        videoFilters.push(`trim=end_frame=${inputFrameCount}`);
+    }
+    else {
+        videoFilters.push(`trim=duration=${originalDuration}`);
+    }
     // Apply speed if specified
     if (speed === 1.0) {
         videoFilters.push("setpts=PTS-STARTPTS");
@@ -79,7 +92,7 @@ const getOutputOption = (audioId, videoId) => {
         "4M", // Reduced buffer size
         "-maxrate",
         "3M", // Reduced from 7M to 3M
-        "-r 30", // Set frame rate
+        `-r ${VIDEO_FPS}`, // Set frame rate
         "-pix_fmt yuv420p", // Set pixel format for better compatibility
         "-c:a aac", // Audio codec
         "-b:a 128k", // Audio bitrate
@@ -284,7 +297,7 @@ const getClampedTransitionDuration = (transitionDuration, prevBeatDuration, curr
     return Math.min(transitionDuration, maxDuration);
 };
 export const getTransitionFrameDurations = (context, index) => {
-    const minFrame = 1 / 30; // 30fpsを想定。最小1フレーム
+    const minFrame = 1 / VIDEO_FPS;
     const beats = context.studio.beats;
     const scriptBeats = context.studio.script.beats;
     const getTransitionDuration = (transition, prevBeatIndex, currentBeatIndex) => {
@@ -320,22 +333,22 @@ export const addSplitAndExtractFrames = (ffmpegContext, videoId, firstDuration,
     if (needFirst) {
         // Create static frame using nullsrc as base for proper framerate/timebase
         // Note: setpts must NOT be used here as it loses framerate metadata needed by xfade
-        ffmpegContext.filterComplex.push(`nullsrc=size=${canvasInfo.width}x${canvasInfo.height}:duration=${firstDuration}:rate=30[${videoId}_first_null]`);
+        ffmpegContext.filterComplex.push(`nullsrc=size=${canvasInfo.width}x${canvasInfo.height}:duration=${firstDuration}:rate=${VIDEO_FPS}[${videoId}_first_null]`);
         ffmpegContext.filterComplex.push(`[${videoId}_first_src]select='eq(n,0)',scale=${canvasInfo.width}:${canvasInfo.height}[${videoId}_first_frame]`);
-        ffmpegContext.filterComplex.push(`[${videoId}_first_null][${videoId}_first_frame]overlay=format=auto,fps=30[${videoId}_first]`);
+        ffmpegContext.filterComplex.push(`[${videoId}_first_null][${videoId}_first_frame]overlay=format=auto,fps=${VIDEO_FPS}[${videoId}_first]`);
     }
     if (needLast) {
         if (isMovie) {
             // Movie beats: extract actual last frame
-            ffmpegContext.filterComplex.push(`nullsrc=size=${canvasInfo.width}x${canvasInfo.height}:duration=${lastDuration}:rate=30[${videoId}_last_null]`);
+            ffmpegContext.filterComplex.push(`nullsrc=size=${canvasInfo.width}x${canvasInfo.height}:duration=${lastDuration}:rate=${VIDEO_FPS}[${videoId}_last_null]`);
             ffmpegContext.filterComplex.push(`[${videoId}_last_src]reverse,select='eq(n,0)',reverse,scale=${canvasInfo.width}:${canvasInfo.height}[${videoId}_last_frame]`);
-            ffmpegContext.filterComplex.push(`[${videoId}_last_null][${videoId}_last_frame]overlay=format=auto,fps=30[${videoId}_last]`);
+            ffmpegContext.filterComplex.push(`[${videoId}_last_null][${videoId}_last_frame]overlay=format=auto,fps=${VIDEO_FPS}[${videoId}_last]`);
         }
         else {
             // Image beats: all frames are identical, so just select one
-            ffmpegContext.filterComplex.push(`nullsrc=size=${canvasInfo.width}x${canvasInfo.height}:duration=${lastDuration}:rate=30[${videoId}_last_null]`);
+            ffmpegContext.filterComplex.push(`nullsrc=size=${canvasInfo.width}x${canvasInfo.height}:duration=${lastDuration}:rate=${VIDEO_FPS}[${videoId}_last_null]`);
             ffmpegContext.filterComplex.push(`[${videoId}_last_src]select='eq(n,0)',scale=${canvasInfo.width}:${canvasInfo.height}[${videoId}_last_frame]`);
-            ffmpegContext.filterComplex.push(`[${videoId}_last_null][${videoId}_last_frame]overlay=format=auto,fps=30[${videoId}_last]`);
+            ffmpegContext.filterComplex.push(`[${videoId}_last_null][${videoId}_last_frame]overlay=format=auto,fps=${VIDEO_FPS}[${videoId}_last]`);
         }
     }
 };
@@ -367,6 +380,7 @@ export const createVideo = async (audioArtifactFilePath, outputVideoPath, contex
     const needsFirstFrame = getNeedFirstFrame(context);
     // Check which beats need _last (for any transition on next beat - they all need previous beat's last frame)
     const needsLastFrame = getNeedLastFrame(context);
+    let cumulativeFrames = 0;
     context.studio.beats.reduce((timestamp, studioBeat, index) => {
         const beat = context.studio.script.beats[index];
         if (beat.image?.type === "voice_over") {
@@ -377,13 +391,19 @@ export const createVideo = async (audioArtifactFilePath, outputVideoPath, contex
         const sourceFile = isTest ? "/test/dummy.mp4" : validateBeatSource(studioBeat, index);
         // The movie duration is bigger in case of voice-over.
         const duration = Math.max(studioBeat.duration + getExtraPadding(context, index), studioBeat.movieDuration ?? 0);
+        // Use cumulative frame tracking to prevent audio-video drift from frame quantization.
+        // trim=duration=X rounds up to the next frame boundary (~0.03s per beat at 30fps),
+        // causing cumulative drift. Instead, compute exact frame counts per beat.
+        const targetEndFrame = Math.round((timestamp + duration) * VIDEO_FPS);
+        const frameCount = targetEndFrame - cumulativeFrames;
+        cumulativeFrames = targetEndFrame;
         const inputIndex = FfmpegContextAddInput(ffmpegContext, sourceFile);
         const isMovie = !!(studioBeat.lipSyncFile ||
             studioBeat.movieFile ||
             MulmoPresentationStyleMethods.getImageType(context.presentationStyle, beat) === "movie");
         const speed = beat.movieParams?.speed ?? 1.0;
         const filters = beat.movieParams?.filters;
-        const { videoId, videoPart } = getVideoPart(inputIndex, isMovie, duration, canvasInfo, getFillOption(context, beat), speed, filters);
+        const { videoId, videoPart } = getVideoPart(inputIndex, isMovie, duration, canvasInfo, getFillOption(context, beat), speed, filters, frameCount);
         ffmpegContext.filterComplex.push(videoPart);
         // for transition
         const needFirst = needsFirstFrame[index]; // This beat has slidein

package/lib/types/provider2agent.d.ts

@@ -155,7 +155,7 @@ export declare const provider2LLMAgent: {
         readonly agentName: "geminiAgent";
         readonly defaultModel: "gemini-2.5-flash";
         readonly max_tokens: 8192;
-        readonly models: readonly ["gemini-3-pro-preview", "gemini-3-flash-preview", "gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.5-flash-lite"];
+        readonly models: readonly ["gemini-3.1-pro-preview", "gemini-3-flash-preview", "gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.5-flash-lite"];
         readonly keyName: "GEMINI_API_KEY";
     };
     readonly groq: {

package/lib/types/provider2agent.js

@@ -289,7 +289,7 @@ export const provider2LLMAgent = {
         agentName: "geminiAgent",
         defaultModel: "gemini-2.5-flash",
         max_tokens: 8192,
-        models: ["gemini-3-pro-preview", "gemini-3-flash-preview", "gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.5-flash-lite"],
+        models: ["gemini-3.1-pro-preview", "gemini-3-flash-preview", "gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.5-flash-lite"],
         keyName: "GEMINI_API_KEY",
     },
     groq: {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mulmocast",
-  "version": "2.4.2",
+  "version": "2.4.4",
   "description": "",
   "type": "module",
   "main": "lib/index.node.js",