fix(canvas): 视频参考协议 - 统一首尾帧与角色资产语义

This commit is contained in:
ddCat committed 2026-08-30 16:52:09 +08:00
1 parent b5f8b66622
commit 37eb2b7e48
15 files changed
+479 -68

No files matched your search

+5 -12
View File
@@ -9,22 +9,22 @@ import type { ReferenceAudio, ReferenceVideo } from "@/types/media";
import type { MiniMaxVideoCreateResponse, MiniMaxVideoTask, RequestOptions, ResolvedAiConfig, VideoGenerationTask, VideoGenerationTaskState } from "./video-contracts";
import type { VideoProviderDeps } from "./video-provider-deps";
import { resolveVideoImageReferences } from "./video-reference-roles";
export async function createMiniMaxVideoTask(deps: VideoProviderDeps, config: ResolvedAiConfig, model: string, prompt: string, references: ReferenceImage[], videoReferences: ReferenceVideo[], audioReferences: ReferenceAudio[], options?: RequestOptions): Promise<VideoGenerationTask> {
const imageUrls = await Promise.all(references.slice(0, 9).map((image) => resolveMiniMaxImageUrl(image)));
const videoUrls = await Promise.all(videoReferences.slice(0, 3).map((video) => resolveMiniMaxMediaUrl(video, "参考视频")));
const audioUrls = await Promise.all(audioReferences.slice(0, 3).map((audio) => resolveMiniMaxMediaUrl(audio, "参考音频")));
const content: Array<Record<string, unknown>> = [{ type: "text", text: prompt.trim() }];
const explicitFrameMode = options?.videoEditOperation !== "reference_to_video" && Boolean(options?.videoStartFrameNodeId || options?.videoEndFrameNodeId);
const imagePlan = resolveVideoImageReferences(references.slice(0, 9), options, { videoCount: videoUrls.length, audioCount: audioUrls.length });
const frameMode = imagePlan.some(({ role }) => role === "first_frame" || role === "last_frame");
imageUrls.forEach((url, index) => {
const reference = references[index];
const role = miniMaxImageRole(reference?.id, options);
if (explicitFrameMode && role === "reference_image") throw new Error("MiniMax 首尾帧模式不能混合未标记的参考图");
const role = imagePlan[index]?.role || "reference_image";
if (frameMode && role === "reference_image") throw new Error("MiniMax 首尾帧模式不能混合未标记的参考图");
content.push({ type: "image_url", image_url: { url }, role });
});
videoUrls.forEach((url) => content.push({ type: "video_url", video_url: { url }, role: "reference_video" }));
audioUrls.forEach((url) => content.push({ type: "audio_url", audio_url: { url }, role: "reference_audio" }));
const frameMode = explicitFrameMode;
const payload = {
model: modelOptionName(model),
content,
@@ -43,13 +43,6 @@ export async function createMiniMaxVideoTask(deps: VideoProviderDeps, config: Re
}
}
function miniMaxImageRole(imageId: string | undefined, options?: RequestOptions) {
if (options?.videoEditOperation === "reference_to_video") return "reference_image";
if (imageId && imageId === options?.videoStartFrameNodeId) return "first_frame";
if (imageId && imageId === options?.videoEndFrameNodeId) return "last_frame";
return "reference_image";
}
export async function pollMiniMaxVideoTask(deps: VideoProviderDeps, config: ResolvedAiConfig, task: VideoGenerationTask, options?: RequestOptions): Promise<VideoGenerationTaskState> {
try {
const response = await deps.transport.get<{ task?: MiniMaxVideoTask }>(miniMaxVideoUrl(config, `/query/video_generation/${encodeURIComponent(task.id)}`), options);