@memberjunction/ai-blackforestlabs

MemberJunction AI provider for Black Forest Labs’ FLUX image generation models. This package implements the BaseImageGenerator interface to provide text-to-image generation, image editing (inpainting/outpainting), and image variations through BFL’s async task-based API.

Architecture

graph TD
    A["FLUXImageGenerator<br/>(Provider)"] -->|extends| B["BaseImageGenerator<br/>(@memberjunction/ai)"]
    A -->|submits tasks| C["BFL REST API"]
    A -->|polls| D["Result Endpoint"]
    D -->|returns| E["Image URL"]
    A -->|downloads| F["Generated Image"]
    B -->|registered via| G["@RegisterClass"]

    style A fill:#7c5295,stroke:#563a6b,color:#fff
    style B fill:#2d6a9f,stroke:#1a4971,color:#fff
    style C fill:#2d8659,stroke:#1a5c3a,color:#fff
    style D fill:#2d8659,stroke:#1a5c3a,color:#fff
    style E fill:#b8762f,stroke:#8a5722,color:#fff
    style F fill:#2d6a9f,stroke:#1a4971,color:#fff
    style G fill:#b8762f,stroke:#8a5722,color:#fff

Async Task Flow

sequenceDiagram
    participant App as Application
    participant FLUX as FLUXImageGenerator
    participant BFL as BFL API

    App->>FLUX: GenerateImage(params)
    FLUX->>BFL: POST /flux-2-pro (submit task)
    BFL-->>FLUX: { id: "task-123", polling_url: "https://api.bfl.ai/v1/get_result?id=task-123", cost: 4 }
    loop Poll polling_url until ready
        FLUX->>BFL: GET polling_url (or /get_result?id=task-123)
        BFL-->>FLUX: { status: "Pending" }
    end
    BFL-->>FLUX: { status: "Ready", cost: 4, result: { sample: "url" } }
    FLUX->>BFL: Download image from signed URL
    BFL-->>FLUX: Image data
    FLUX-->>App: ImageGenerationResult (with usage.cost in USD)

Features

• Text-to-Image Generation: Generate photorealistic images using FLUX.2 Pro, FLUX 1.1 Pro, and other FLUX models
• Image Editing: Inpainting and outpainting via FLUX Fill model with mask support
• Image Variations: Create variations using FLUX Kontext Pro for image-to-image generation
• Async Task Polling: Configurable polling with adjustable interval and timeout
• Multiple Models: Support for FLUX.2 Pro (32B), FLUX 1.1 Pro, FLUX Dev, and FLUX Ultra
• Seed Reproducibility: Deterministic generation with seed parameter support
• Resolution Control: Custom width/height and aspect ratio settings up to 4MP

Installation

npm install @memberjunction/ai-blackforestlabs

Usage

import { FLUXImageGenerator } from '@memberjunction/ai-blackforestlabs';

const generator = new FLUXImageGenerator('your-bfl-api-key');

const result = await generator.GenerateImage({
    prompt: 'A serene mountain landscape at sunset',
    model: 'flux-2-pro',
    size: '1536x1024'
});

if (result.success) {
    const image = result.images[0];
    console.log('Image generated, base64 length:', image.base64.length);
}

Configure Polling

generator.setPollingConfig({
    maxWaitTime: 180000,  // 3 minutes
    pollInterval: 3000     // 3 seconds
});

Supported Models

Model	ID	Modality	Description
FLUX 3 Image	flux-3-image	Image	Up to 4K resolution, 16 aspect ratios, and up to 10 reference images
FLUX 3 Video	flux-3-video	Video	Text-to-video and image-to-video (5–20s), continuation via v2v (5–15s); API supports hd..uhd (driver maps Quality up to qhd) with optional native audio; driver refuses VideoToVideo and measures MP4 bytes for C2PA
FLUX.2 Pro	flux-2-pro	Image	32B parameter, photorealistic 4MP images
FLUX 1.1 Pro	flux-1.1-pro	Image	Production-ready, high-volume workloads
FLUX.1 Dev	flux-dev	Image	Open-weights 12B for development
FLUX Ultra	flux-ultra	Image	Ultra-high quality

Class Registration

• Image generator: registered as FLUXImageGenerator via @RegisterClass(BaseImageGenerator, 'FLUXImageGenerator').
• Video generator: registered as FLUX3VideoGenerator via @RegisterClass(BaseVideoGenerator, 'FLUX3VideoGenerator').

Verified against BFL docs

• Verification date: 2026-10-16
• Documentation URLs read:
  • https://docs.bfl.ai/flux_3/flux3_overview (flux_3/flux3_overview.md)
  • https://docs.bfl.ai/flux_3/flux3_video (flux_3/flux3_video.md)
  • https://docs.bfl.ai/flux_3/flux3_image (flux_3/flux3_image.md)
  • https://docs.bfl.ai/flux_3/flux3_image_generate (flux_3/flux3_image_generate.md)
  • https://docs.bfl.ai/flux_tools/flux_video_edit (flux_tools/flux_video_edit.md)
  • https://docs.bfl.ai/cookbook/video_quickstart (cookbook/video_quickstart.md)
  • https://docs.bfl.ai/cookbook/video_edit_recast_continue (cookbook/video_edit_recast_continue.md)
  • https://docs.bfl.ai/api_integration/integration_guidelines (api_integration/integration_guidelines.md)
  • https://docs.bfl.ai/api-reference/utility/generate-a-video-with-flux-3 (api-reference/utility/generate-a-video-with-flux-3.md)
  • https://docs.bfl.ai/api-reference/utility/generate-an-image-with-flux-3 (api-reference/utility/generate-an-image-with-flux-3.md)
  • https://docs.bfl.ai/api-reference/utility/get-result (api-reference/utility/get-result.md)
  • https://docs.bfl.ai/flux_2/flux2_text_to_image (flux_2/flux2_text_to_image.md)
  • https://bfl.ai/pricing (1 credit = $0.01 USD)
• Endpoints & Authentication:
  • Base URLs: https://api.bfl.ai/v1 (global), https://api.eu.bfl.ai/v1, https://api.us.bfl.ai/v1
  • Auth header on submit and poll: x-key: <BFL_API_KEY> (never sent when downloading signed delivery URLs on delivery.*.bfl.ai)
  • FLUX 3 Video submit: POST /v1/flux-3-video
  • FLUX 3 Image submit: POST /v1/flux-3-image
  • FLUX.2 Pro submit: POST /v1/flux-2-pro
  • Result polling: GET <polling_url> from submit response (validated to be https: and hostname ending in .bfl.ai), or fallback GET /v1/get_result?id=<id> when polling_url is omitted
• FLUX 3 Video (Flux3VideoInputsBody, additionalProperties: false) parameters & limits:
  • mode: 't2v' | 'i2v' | 'v2v' (required; 'v2v' on /v1/flux-3-video is video continuation via start_video, whereas video editing is a separate endpoint /v1/flux-tools/video-edit-v1; FLUX3VideoGenerator refuses Mode: 'VideoToVideo')
  • prompt: string (required)
  • resolution: 'hd' | 'fhd' | 'qhd' | 'uhd' (API schema default 'hd'; driver maps Quality: 'draft' -> 'hd' + draft: true, 'standard' -> 'fhd', 'high' -> 'qhd')
  • aspect_ratio: 'auto' | '21:9' | '2:1' | '16:9' | '4:3' | '1:1' | '3:4' | '9:16' | '9:21' (API schema default 'auto'; driver requires one of the 8 fixed ratios so pixel dimensions are known; '4:5' is not supported)
  • duration: 'auto' or integer 5..20 seconds for 't2v' and 'i2v'; 5..15 seconds for 'v2v' (API schema default 'auto'; driver requires an integer in 5..20 for 't2v'/'i2v' or 5..15 for 'v2v')
  • generate_audio: boolean (BFL API defaults to true; FLUX3VideoGenerator always sends generate_audio: params.WantAudio === true so omitted WantAudio never bills for audio)
  • draft: boolean (optional; set to true for 'draft' quality tier)
  • keyframes: Array<[number, string]> | string[] (1..10 items; required in 'i2v' mode; timestamped [seconds, image] pairs with 0 <= seconds <= duration or plain base64/HTTPS URL strings)
  • start_video: string ('v2v' continuation mode only; base64 or HTTPS URL; mp4; ≤50 MB; ≤15s per cookbook/video_edit_recast_continue.md)
  • safety_tolerance: integer 0..7 (default 2)
  • user, version: optional strings in the t2v/i2v/v2v input schemas (webhook_url and webhook_secret are not in the t2v/i2v/v2v input schemas)
  • Does not accept seed, fps, width, or height.
• FLUX 3 Video 8×4 Dimension Lookup Table (FLUX3_VIDEO_DIMENSIONS, width × height):

Aspect Ratio	hd (720p)	fhd (1080p)	qhd (1440p)	uhd (4K)
21:9	1440×608	2176×928	2912×1248	4352×1856
2:1	1344×672	2016×1024	2720×1376	4032×2048
16:9	1280×704	1920×1088	2560×1440	3840×2176
4:3	1088×800	1632×1216	2176×1632	3264×2432
1:1	960×960	1440×1440	1920×1920	2880×2880
3:4	800×1088	1216×1632	1632×2176	2432×3264
9:16	704×1280	1088×1920	1440×2560	2176×3840
9:21	608×1440	928×2176	1248×2912	1856×4352

• FLUX 3 Image (Flux3ImageInputs, additionalProperties: false) parameters & limits:
  • prompt: string (required)
  • images: string[] | null (optional; up to 10 reference images)
  • aspect_ratio: 'auto' | '21:9' | '2:1' | '16:9' | '3:2' | '7:5' | '4:3' | '5:4' | '1:1' | '4:5' | '3:4' | '5:7' | '2:3' | '9:16' | '1:2' | '9:21' (API schema default 'auto'; driver omits aspect_ratio when no dimensions or aspect ratio are specified so BFL uses 'auto')
  • resolution: '768sq' | '1k' | '1.5k' | '2k' | '4k' (default '1k')
  • grounding: boolean (BFL API defaults to true; FLUXImageGenerator always sets grounding: false so image generation never makes unsolicited web searches)
  • safety_tolerance: integer 0..7 (default 2)
  • Does not accept width, height, seed, or output_format.
• Polling Statuses (StatusResponse — 9 values):
  • Active: 'Pending', 'Reasoning', 'Generating'
  • Success: 'Ready' (result.sample contains the signed delivery URL)
  • Terminal failures: 'Error', 'Request Moderated', 'Content Moderated', 'No Card Generated' (surfaces details['Moderation Reasons'] when present), 'Task not found'
• Pricing Status (bfl.ai/pricing, 1 credit = $0.01 USD):
  • Submit responses include initial cost (credits) and /v1/get_result includes settled cost (credits) on 'Ready'. Both drivers convert (pollResult.cost ?? submitResult.cost) / 100 to USD on ModelUsage.
  • Published FLUX 3 Video credits per output second on bfl.ai/pricing:
    • Draft 720p (hd): 6 credits/s ($0.06/s)
    • Standard t2v / i2v: 720p (hd) 15 credits/s ($0.15/s), 1080p (fhd) 24 credits/s ($0.24/s), 1440p (qhd) 28 credits/s ($0.28/s), 2160p (uhd) 41 credits/s ($0.41/s)
    • Video-to-video (v2v continuation): 720p (hd) 21 credits/s ($0.21/s), 1080p (fhd) 27 credits/s ($0.27/s), 1440p (qhd) 31 credits/s ($0.31/s), 2160p (uhd) 45 credits/s ($0.45/s)
  • Published FLUX 3 Image rates per image on bfl.ai/pricing: 768×768 $0.041, 1K $0.048, 1.5K $0.07, 2K $0.10, 4K $0.607 (reference images and prompts are not charged extra).
• Delivery URL Expiry Conflict:
  • BFL’s documentation states three different expiry windows for signed result.sample delivery URLs: 10 minutes (integration guide), ~1 hour (FLUX 3 text-to-image and Video Edit pages), and ~2 hours (video quickstart and FLUX 3 overview). Both drivers download bytes immediately upon 'Ready' and never persist result.sample.
• C2PA Provenance:
  • BFL’s help center documents C2PA embedding for API images only (nothing in BFL’s docs states C2PA for video). FLUX3VideoGenerator measures the downloaded MP4 bytes at runtime with HasC2PAManifest(bytes) (top-level ISO-BMFF uuid box with extended type d8fec3d6-1b0e-483c-9297-5828877ec481, C2PA 2.2 §A.5.1.1) and sets result.Provenance = { C2PA: HasC2PAManifest(bytes) }.

Opt-in Live Smoke Test (Run Only With Explicit Approval)

Automated test suites in CI and local builds never make live network calls to BFL. To run a live end-to-end smoke check (only with explicit approval from Colin):

• Prerequisites:
  1. metadata/prompts/.default-generative-video-prompt.json (VS-GEN-2) and the active FLUX 3 Video model + Per Second price unit (VS-GEN-4) are synced to the target SQL Server database.
  2. Standard MJ database environment variables (DB_HOST, DB_PORT, DB_DATABASE, DB_USERNAME, DB_PASSWORD) and BFL_API_KEY are set.
  3. Run from packages/MJServer (after building @memberjunction/ai-blackforestlabs and @memberjunction/ai-prompts).
• Expected spend: 1 draft 5-second 720p (hd) clip ≈ 30 credits ($0.30 USD) at bfl.ai/pricing’s draft rate of 6 credits/s.

cd packages/MJServer
node --input-type=module -e '
import { writeFileSync } from "node:fs";
import sql from "mssql";
import { setupSQLServerClient, SQLServerProviderConfigData, UserCache } from "@memberjunction/sqlserver-dataprovider";
import { RunView } from "@memberjunction/core";
import { AIEngine } from "@memberjunction/aiengine";
import { AIVideoRunner } from "@memberjunction/ai-prompts";
import "../AI/Providers/BlackForestLabs/dist/index.js";

const pool = new sql.ConnectionPool({
    server: process.env.DB_HOST ?? "localhost",
    port: Number(process.env.DB_PORT ?? 1433),
    database: process.env.DB_DATABASE,
    user: process.env.DB_USERNAME,
    password: process.env.DB_PASSWORD,
    options: { encrypt: true, trustServerCertificate: true },
});
await pool.connect();
await setupSQLServerClient(new SQLServerProviderConfigData(pool, process.env.MJ_CORE_SCHEMA ?? "__mj"));

const user = UserCache.Instance.Users[0];
await AIEngine.Instance.Config(false, user);

const runner = new AIVideoRunner();
const result = await runner.RunGenerateVideo({
    Prompt: "A calm ocean wave rolling onto a sandy beach at sunrise",
    Mode: "TextToVideo",
    DurationSeconds: 5,
    AspectRatio: "16:9",
    Quality: "draft",
    WantAudio: false,
    ContextUser: user,
    CredentialScope: "RuntimeOnly",
    APIKeys: [{ driverClass: "FLUX3VideoGenerator", apiKey: process.env.BFL_API_KEY }],
});

if (!result.Success || !result.Video?.Data?.length) {
    console.error("Smoke test failed:", result.ErrorMessage);
    process.exit(1);
}

writeFileSync("/tmp/flux3-smoke.mp4", result.Video.Data);
console.log("Saved /tmp/flux3-smoke.mp4", {
    bytes: result.Video.Data.length,
    width: result.Video.Width,
    height: result.Video.Height,
    durationSeconds: result.Video.DurationSeconds,
    c2pa: result.C2PA,
    promptRunId: result.PromptRunID,
});

const rv = new RunView();
const runRes = await rv.RunView({
    EntityName: "MJ: AI Prompt Runs",
    ExtraFilter: `ID = '\''${result.PromptRunID}'\''`,
    Fields: ["ID", "Success", "Cost", "CostCurrency", "TokensUsedRollup"],
    ResultType: "simple",
}, user);
console.log("PromptRun row:", runRes.Results[0]);
await pool.close();
process.exit(0);
'

Then open /tmp/flux3-smoke.mp4 to verify playback and confirm PromptRun row has a non-null Cost (~0.30 USD).

Dependencies

• @memberjunction/ai - Core AI abstractions (BaseImageGenerator, BaseVideoGenerator, GeneratedImage)
• @memberjunction/global - Class registration