Video Generation

Nội dung này hiện chưa có sẵn bằng ngôn ngữ của bạn.

The video generation subsystem coordinates asynchronous AI video generation across external providers and custom endpoints for the /generate video command and the generate_video built-in tool.

Command and tool entry points

Video generation exposes two user-facing surfaces:

Slash command (src/commands/generate/video.ts): Opens a modal collecting prompt text, aspect ratio, duration (seconds), optional frame rate (FPS), and an optional reference image attachment.
Built-in tool (src/tools/functionCalls/generateVideoTool.ts): Allows models to invoke generate_video with a prompt, target duration, resolution (480p, 720p, 1080p), optional starting image (media_id), and looping controls. Tool progress notices display the active model codename, prompt excerpt, reference image usage, and elapsed execution time.

Administrative controls manage model selection and quotas:

Model selection: /config > Models > Switch Models (src/utils/discord/interactions/configModelRoutes.ts) binds server_model_configs.video_model_id.
Quota administration: /moderation (Quotas page) and /quota reset (src/commands/quota/reset/global.ts, src/commands/quota/reset/user.ts) manage user and server-wide quotas.

Asynchronous job lifecycle and providers

Provider implementations are resolved via resolveProviderFeatureImplementation in src/utils/provider/providerInfoRegistry.ts:

Google Gemini Omni: src/providers/google/googleVideoGeneration.ts
OpenRouter: src/providers/openrouter/openrouterVideoGeneration.ts
Z.ai: src/providers/zai/zaiVideoGeneration.ts
Custom ComfyUI endpoints: src/providers/custom/customEndpointDispatcher.ts via generateCustomVideoViaEndpoint

Video adapters submit remote work and poll for completion before downloading MP4 data. Polling budgets belong to each adapter: Google and Z.ai use roughly five minutes of polling intervals; OpenRouter uses roughly ten. Network request time adds to those intervals.

OpenRouter asynchronous API

OpenRouter video generation submits jobs to POST /api/v1/videos. The adapter polls the returned polling_url until the job reaches a terminal status, then downloads the binary video from unsigned_urls (falling back to /api/v1/videos/{jobId}/content?index=0).

To protect credentials, relative polling URLs are resolved against https://openrouter.ai, and authenticated polling is restricted to that origin. A dedicated video model cache (src/utils/cache/openrouterVideoModelCache.ts) queries GET /api/v1/videos/models to validate supported durations, resolutions, aspect ratios, and frame capabilities before submitting jobs.

The command and tool default Grok Imagine Video 1.5 Lite to 480p. The tool preserves an explicit resolution choice; openrouterVideoRequest.ts normalizes it against the model’s supported values.

Image-to-video requests pass reference images via frame_images with frame_type: "first_frame". Setting loop mode supplies the same image as last_frame when the model supports it. If a model lacks first-frame or last-frame capabilities, the request fails before submitting a paid job.

Transport bypass for OpenRouter

The OpenRouter video adapter uses an external HTTP process to work around HTML responses observed with Bun’s transport at the provider’s edge. This is an adapter-specific compatibility measure; it does not establish a permanent rule about the provider’s TLS filtering.

To preserve connectivity, src/providers/openrouter/openrouterVideoGeneration.ts delegates requests to externalHttpRequest(), which spawns an external process with standard TLS fingerprints:

Windows: PowerShell 7 (pwsh) with Invoke-WebRequest, using .NET Schannel TLS with HTTP/2 negotiation. Request data passes via stdin as JSON, and binary output returns base64-encoded.
Linux and Docker: curl with HTTP/2 via nghttp2 (--proto =https, -H "Expect:"). Headers, including the bearer token, and the request body travel through stdin as a curl config (-K -), so neither appears in the process table.

Google and Z.ai adapters use the native transport.

State and quota coordination

Video generation persists configuration in server-scoped tables:

server_capabilities_configs.videogen_enabled: Feature gate for the command and tool.
server_model_configs.video_model_id: Active video model foreign key.
saved_provider_configs.video_model_id: Preserved model slot when switching provider profiles.

Quota tracking and delivery invariant

Video quotas use video_quota_configs, video_quotas, and video_serverwide_quotas. Both user and server limits default to unlimited (0); a configured server limit uses a 365-day default reset period.

Server-funded generation reserves quota after the modal is validated and before submission. Personal-provider selections bypass server quota accounting. The reservation is a conditional increment on both counters in one transaction, so concurrent requests cannot all pass the same preflight. A failed, cancelled, or undelivered generation releases the unit it reserved; a delivered video keeps it. Tool cancellation checks the turn signal before delivery, but a submitted provider job may still run or incur charges. When the counters cannot be read or claimed, the command and tool refuse with a localized quota-unavailable notice instead of generating uncounted. The quota reset boundary belongs to the reservation: a release matches the reserved day and serverwide period, so a job that finishes after a reset leaves the new window untouched.

Discord delivery constraints

Generated videos must satisfy Discord attachment limits:

File size limit: Command and tool enforce a local 25 MiB ceiling (DISCORD_FILE_SIZE_LIMIT). This fixed ceiling does not represent every Discord account or guild’s current upload allowance.
Format: Providers return binary MP4 video.
Components V2 presentation: src/utils/discord/generatedVideoMessage.ts wraps the attachment in a Components V2 message. A MediaGallery item references the attachment:// file so Discord renders its inline player, and a TextDisplay component places a localized generation time notice (“Generated in Xs”) below the player. If Components V2 rendering fails, delivery falls back to a standard attachment-only message.

Source pointers

src/commands/generate/video.ts: Slash command modal flow and input normalization.
src/tools/functionCalls/generateVideoTool.ts: Built-in tool execution and progress notice dispatch.
src/providers/openrouter/openrouterVideoGeneration.ts: OpenRouter async polling and external HTTP process dispatcher.
src/providers/google/googleVideoGeneration.ts, src/providers/zai/zaiVideoGeneration.ts: Native provider video adapters.
src/utils/quota/videoQuotaManager.ts: quota validation, reservation, and release logic.
src/utils/discord/generatedVideoMessage.ts: Components V2 media gallery message formatting.