FLUX 3 Image: Custom Output Sizes and Better Edits
POST /v1/flux-3-image has two new size options: width and height, and resolution: "match_input" for edits. Edits with one reference image are also more accurate.Key Features
- Exact size: send
widthandheightfor any size from 256px per side up to 16.8MP (4k), at aspect ratios up to 8:1. With reference images, the limit is 4MP.widthandheightoverrideresolutionandaspect_ratio - Same size as input image: with
resolution: "match_input", an edit is the same size as the first reference image, up to 4MP. Ifaspect_ratioor your prompt sets an aspect ratio, the output uses it at a standard size - Better edits: edits with one reference image follow the instruction more closely, especially whole-image changes such as weather, relighting, color grading, and restoration, plus text edits and camera moves. No request changes needed
- No cropping: with
aspect_ratioatauto, edits with one reference image use the whole image - Backward compatible: requests without the new fields work as before and return
1k
Pricing
Prices are unchanged. Custom andmatch_input sizes are billed at the smallest resolution with at least as many pixels. Anything above 2k (about 4MP) is billed as 4k.Endpoint
FLUX 3 Image API referenceFLUX 3 Image Klein
FLUX 3 Image Klein is a 5.6B parameter version of FLUX 3 Image with open weights. It supports the same features: text to image, image editing, multi reference, and bounding boxes. Run it through the API, download the weights from Hugging Face, or load it in ComfyUI.Key Features
- Same request as FLUX 3 Image:
prompt,images,aspect_ratio,resolution, and the bounding box format are unchanged. A FLUX 3 Image request runs on Klein by switching the endpoint - Two variants: Klein (Distilled) runs 16 steps at CFG 1 for fast inference. Klein Base runs 50 steps at CFG 3.5 and is the model for fine-tuning and LoRA training
- Open weights: both variants are on Hugging Face under the FLUX Kommunity License and run on consumer GPUs, down to an RTX 3090
- Up to 4K output:
resolutionfrom768sqto4k, with up to ten reference images per request - LoRA training: train on Klein Base with AI Toolkit (
arch: flux3) or Diffusers. The trained LoRA also loads into Klein (Distilled)
Pricing
Per image, byresolution. Text to image and image editing cost the same.Endpoint
FLUX 3 Image Klein API referenceWeights
- FLUX 3 Image Klein (Distilled)
- FLUX 3 Image Klein Base
- ComfyUI weights and the ComfyUI Klein workflow
FLUX 3 Image
FLUX 3 Image gives you maximum control over every pixel. Generate an image from a prompt, edit specific parts of an image, and combine up to ten references, all through one endpoint.Key Features
- Bounding boxes in the prompt: place every element of a layout, such as a headline, a face in a crowd, or each panel of a grid. Add a JSON list of boxes to the end of
prompt; there is no separate parameter - Pixel-exact local edits: mark the element to change and keep the rest of the image as it was. Recolor, replace, move, resize, or remove several elements in one request
- Up to ten references: name each image by its position in the prompt to restyle an image or combine references into a new composition
- Up to 4K output:
resolutiongoes from768sqfor fast previews to4k(about 16 megapixels), at 15 aspect ratios - One endpoint: the prompt decides whether to generate or edit. There is no
modefield
Pricing
Per image, byresolution:Endpoint
FLUX 3 Image API referenceStart with the FLUX 3 Image overview, see the box format in FLUX 3 Image Editing, or follow the Bounding boxes tutorial.FLUX 3 Video — 2K and 4K Output
resolution on POST /v1/flux-3-video now takes qhd and uhd alongside hd and fhd, so a 2K or 4K clip comes back from a single request instead of a generation followed by a separate upscale call.Key Features
qhd— 2560 × 1440 at 16:9uhd— 3840 × 2176 at 16:9- Same generation, larger canvas — every class above
hdruns the same generation and differs only in the canvas the video upsampler renders, so the shot does not change with the resolution you pick - Available on
draft_enhancetoo — commit a draft straight to 2K or 4K. Drafts themselves still render athd
Pricing
Per second of output, in addition to the existinghd and fhd rates:Endpoint
POST /v1/flux-3-videoDelivered dimensions for every aspect ratio are in the FLUX 3 video documentation, and full rates on the pricing page.Auto Top-Up for Credits
You can now recharge your credit balance automatically. Set a threshold and a target on the Credits page, and whenever your balance falls below the threshold, your saved card is charged to bring it back up to the target.- When balance falls below — the balance that triggers a recharge, minimum $5
- Recharge balance to — the balance after a recharge, at least $10 above the threshold
FLUX Video Upscale — 1080p, 2K, and 4K Video Super-Resolution
FLUX Video Upscale brings FLUX 3 powered super-resolution to video. Send a clip of up to 20 seconds and get it back at up to 4K, with the original audio track preserved.Key Features
- 1.5×–3× upscaling —
upscale_factorscales relative to the source resolution; the output keeps the source aspect ratio, capped at about 14.4 megapixels per frame - Two modes —
creativity: 0(precise) preserves the source exactly and sharpens it, ideal for faces, products, and brand assets;creativity: 1(creative, default) restores and invents fine detail for generated footage, textures, and scenery - Optional
prompt— describe the clip’s content to steer the enhanced detail in creative mode - Source limits — up to 20 seconds, 50MB, and 2560 × 1440; rejected clips are not charged
- Pay for delivered output only — $0.07 (precise) / $0.10 (creative) per megapixel-second
Endpoint
POST /v1/flux-tools/video-upscale-v1See the FLUX Video Upscale documentation to get started, and the pricing page for per-resolution rates.FLUX MCP Server — Video Generation
The FLUX MCP server now generates video with FLUX 3, directly from Claude, Cursor, Codex, Windsurf, or any MCP-compatible client.Key Features
generate_video— Text to video, image to video from keyframes, and video continuation, with a prompt plus optionalduration(5–20 seconds),aspect_ratio, andresolution. Up to 4 clips per call, audio on by defaultenhance_video— Re-render a finished draft at full quality with the same composition, seed, and prompt plan: the commit step of the draft-then-enhance loop
FLUX 3 Video Agent Skills
Six new Agent Skills teach coding agents like Claude Code, Cursor, and Windsurf how to direct FLUX 3 video generations.flux-3-video— Routes a request to the skills it needsflux-3-prompt-doctor— Catches the decisions that change the payload before anything is generatedflux-3-cinematic-inserts— Shot craft for text-to-videoflux-3-keyframes-continuation— Building from supplied images or extending an existing clipflux-3-audio-dialogue— Ambience, effects, and speechflux-3-generate— Submitting, polling, drafts, and downloads
FLUX 3 — Video With Synchronized Audio
FLUX 3 is our first video model: one model trained across image, video, and audio, generating video with synchronized sound. Available now as a preview.Key Features
- Up to 20 seconds at FHD (1920 × 1088 for 16:9), 24 fps, in a single request
- Synchronized audio — multilingual speech with strong lipsync, plus effects and ambience, generated with the frames
- Four modes on one endpoint — text to video (
t2v), image to video with 1–10 pinned keyframes (i2v), video continuation (v2v), and draft enhance (draft_enhance) - Multiple scenes and camera angles in one generation, with accurate in-scene text and typography
Endpoint
POST /v1/flux-3-videoThe final polling response includes cost, the settled cost in credits for the request.See the FLUX 3 documentation to get started, and the pricing page for per-second rates. Video editing and Omni Reference with images and videos will be available soon.Announcing FLUX 3
FLUX 3 is our new multimodal foundation model: one model jointly trained across image, video, and audio, built toward real-world visual intelligence. Read the announcement.Capabilities
- Video with native audio — up to 20 seconds per generation, from text, images, or a reference clip, with multilingual dialogue
- Keyframe control — pin images at chosen moments for controlled transitions
- Broad stylistic range — from candid footage to animation and cinematics, with strong typography and animated designs
- Image synthesis and editing — from the same underlying model
Rollout
Capabilities become available in phases, each after an early access period: video generation and editing first, action prediction with selected partners, image synthesis and editing, and open-weight access to a multimodal backbone.FLUX Virtual Try-On v2 — Sharper Face Preservation
vto-v2 is a new version of FLUX Virtual Try-On focused on keeping the person’s identity intact, with additional gains in garment fidelity.Key Features
- Improved face preservation — markedly better retention of the model’s identity and facial features in the try-on result
- Better garment detail — minor improvements to garment textures, prints, and trims
- Higher-resolution output — inputs up to 4 Megapixels are used as-is, and the result follows the model image’s resolution (up from 2 Megapixels in
vto-v1) - Drop-in migration — identical request and response format to
vto-v1; just swap the endpoint path
Endpoint
POST /v1/flux-tools/vto-v2vto-v2 is now the recommended version. vto-v1 remains available, and we recommend migrating to v2 — see the FLUX Virtual Try-On documentation.FLUX Deblur — Sharpen Blurry Images
A new FLUX Tools endpoint that takes a single blurry image and returns a sharper version of the same scene — no prompt, no mask.Key Features
- Faithful restoration — Reconstructs fine detail while preserving character identity, colors, objects, lighting, and composition
- No prompt required — Send the image; the model sharpens the whole frame automatically
- Up to 4MP input — As base64 or an image URL
- Built for volume — Low-latency and low-cost, suited to interactive and high-throughput workflows
Endpoint
POST /v1/flux-tools/deblur-v1Get started in the FLUX Deblur documentation.Moderation Responses Now Include Reasons
When a request or result is moderated, the polling response’sdetails field now includes a Moderation Reasons array naming the categories that triggered the filter (for example Violence or Sexual Content). Use it to adjust the request or surface a message to your users.This applies to both Request Moderated (input flagged at submission) and Content Moderated (generated output flagged). See the Errors reference for the response format.FLUX Outpainting — New Fast Mode
FLUX Outpainting now supports amode parameter that trades quality for speed.Key Features
mode: "fast"— Significantly faster extension, well-suited for landscapes, backgrounds, textures, and products. Typically lower cost.mode: "high"(default) — Unchanged highest-fidelity path; recommended when fine detail, prompt adherence, or consistency with complex source content matters.- Same contract — Both modes share the same request fields and response format; omit
modeto keep the existinghighbehavior.
Endpoint
POST /v1/flux-tools/outpainting-v1See Choosing a mode in the FLUX Outpainting documentation.FLUX Virtual Try-On — See Any Garment On Any Person
A new FLUX Tools endpoint that applies clothing to a person image while preserving their face and pose — built for interactive, low-latency applications.Key Features
- Flexible references — Single garments, multi-garment outfits, and model-to-model references
- Packshot or on-model — Accepts both flat-lay garment images and on-model imagery
- Low latency — Optimized for fitting rooms, social filters, and interactive experiences
- Up to 2MP input — Larger inputs auto-downscale to ~1MP
Endpoint
POST /v1/flux-tools/vto-v1Get started in the FLUX Virtual Try-On documentation, or try the public demo.FLUX Erase — Remove Anything, Leave No Trace
A new FLUX Tools endpoint for clean, prompt-free object removal. Provide an image and a binary mask; the model removes the masked content and reconstructs the scene coherently.Key Features
- Broad removal scope — Objects, people, text, watermarks, and complex compositions
- Coherent reconstruction — Automatically matches lighting, texture, and background
- Optional edge expansion — Cleaner results on soft-edged subjects
- No prompt required — Just an image and a binary mask
Endpoint
POST /v1/flux-tools/erase-v1Get started in the FLUX Erase documentation, try the public demo, or read the announcement.FLUX Outpainting — Extend Any Image in Any Direction
A new FLUX Tools endpoint that extends an image beyond its borders in a single call — no prompt required.Key Features
- Natural scene extension — Carries lighting, texture, depth, and composition past the original frame
- Flexible canvas control — Direct width / height and reference placement via
reference_offset_x/reference_offset_y - Up to 4MP output — Production-ready resolutions for banners, social formats, and aspect-ratio changes
- No prompt needed — The model continues the existing scene on its own
Endpoint
POST /v1/flux-tools/outpainting-v1Get started in the FLUX Outpainting documentation, try the public demo, or read the announcement.FLUX MCP Server
The official FLUX MCP server is live athttps://mcp.bfl.ai — generate, edit, vary, and browse FLUX images directly inside Claude, Cursor, Codex, Windsurf, or any MCP-compatible client.Key Features
- OAuth sign-in — Connect with your BFL account; no API keys to paste into the conversation
- All FLUX.2 models — Pro, Max, Klein 4B/9B, and Flex, plus the FLUX Tools
- Full creative loop — Generate options in parallel, edit attached images through prompts, branch into variations from any result, and browse your generation history
FLUX.2 LoRA Inference — Public Beta
Run your own trained LoRAs on FLUX.2 [klein] through the BFL API. Train with the tools of your choice (AI-Toolkit, Diffusers, …), upload once in the Dashboard, then serve inference through a managed-finetuned endpoint — no GPUs to provision. This is a new self-serve path, separate from the FLUX.1-era Finetuning API deprecated in October 2025.Key Features
- Bring your own LoRA — Upload trained LoRAs to the Dashboard, where they appear as Finetunes
- Same request shape — Each
-finetunedendpoint matches its base endpoint, with two added parameters:finetune_idandfinetune_strength - Six endpoints — Covering FLUX.2 [klein] 4B, 9B, and Base variants in FP8, plus a BF16 option
- Base-endpoint pricing during beta — Fine-tuned requests are billed at the same rate as the underlying base endpoint
Endpoints
POST /v1/{base_model}-finetuned, e.g. /v1/flux-2-klein-9b-kv-finetunedSee the LoRA inference guide for the full endpoint list and a worked example.FLUX.2 [pro] Speed Upgrade & Preview Endpoint
FLUX.2 [pro] is now approximately 2x faster across both text-to-image and image editing — with no quality loss and no price change.- ~2x faster generation — Across text-to-image and image editing workloads
- No quality loss — On par with the original across all benchmarks
- No price change — Faster generation at the same price
- Same API contract — Drop-in upgrade, no code changes required
New: flux-2-pro-preview Endpoint
We’re introducing flux-2-pro-preview, where our latest FLUX.2 [pro] improvements land first. This speed upgrade is available now on flux-2-pro-preview, and the endpoint will continue to be updated as we ship new capabilities.The existing flux-2-pro endpoint remains unchanged as a fixed snapshot for workflows that require reproducibility.Learn more in the Preview Endpoint documentation.Agent Skills for FLUX
FLUX Agent Skills teach coding agents like Claude Code, Cursor, and Windsurf how to write FLUX API code and direct FLUX generations — prompting patterns, model selection, async generation, rate-limit handling, and webhook verification.Install from the black-forest-labs/skills repository; the skills follow the open agentskills.io specification. See the Agent Skills documentation to get started.FLUX.2 [klein] Launch
We’re excited to announce FLUX.2 [klein] — our fastest model family with sub-second inference.Key Features
- Sub-second generation — Real-time image generation for interactive applications
- Two variants — 4B from ($0.014 / image) and 9B from ($0.015 / image) models
- Runs on consumer hardware — As little as 13GB VRAM required
- Open weights available — 4B under Apache 2.0, 9B under FLUX Non-Commercial License
API Endpoints
flux-2-klein-4b— Fastest, ideal for high-volume and local deploymentflux-2-klein-9b— Balanced quality and speed with 8B Qwen3 text embedder
FLUX.2 [max] Launch
Introducing FLUX.2 [max] — our highest quality model with groundbreaking capabilities.Key Features
- Maximum quality — Highest editing consistency, strongest prompt following, and faithful style representation
- Grounding search — Generate images grounded in real-time information from the web
- Multi-reference editing — Combine up to 10 input images while maintaining identity
- Vast world knowledge — Create visuals of current events, real-time weather, or historical moments
Grounding Search Examples
FLUX.2 [max] can search the web to visualize:- Yesterday’s football scores
- Real-time weather in any city
- Recent news events
- Historical moments from specific dates and locations
Organizations & Projects
We’ve introduced a new way to organize your teams and work with Organizations and Projects.What’s New
- Organizations - Create organizations to group your projects and team members. Each organization has its own billing, credits, and member management.
- Projects - Organize your work into projects with dedicated API keys, usage tracking, and spending limits.
- Role-Based Access Control - Invite team members with granular permissions: Administrator, Developer, Billing Manager, or Viewer.
- Project-Scoped API Keys - API keys are now tied to specific projects for better security and tracking.
- Spending Limits - Set daily, weekly, and monthly credit limits per project.
- Activity Logging - Full audit trail of all organizational activity.
For Existing Users
Your account has been automatically migrated:- A “Default” organization was created with your existing credits
- A “Default” project contains your existing API keys
- Your API keys continue to work - no changes required
FLUX.2 [pro] and [flex] Launch
Introducing the next generation of FLUX models — FLUX.2 [pro] and FLUX.2 [flex].FLUX.2 [pro]
Production-grade image generation at scale.- Top performance at affordable price — High quality generation from $0.03/MP
- Fast turnaround — Optimized for high-volume applications
- Multi-reference editing — Combine up to 8 input images (10 in playground)
- Photorealistic output — Closes the gap between generated and real imagery
FLUX.2 [flex]
Fine-grained control for precise results.- Specialized for typography — Best-in-class text rendering
- Adjustable parameters — Control inference steps (up to 50) and guidance scale (1.5–10)
- Detail preservation — Maintains small details in complex scenes
- Production workflows — Ideal for designs requiring exact specifications
Key Capabilities
Both models support:- Exact color control — Specify brand colors via hex codes with precision matching
- Structured prompting — JSON-style prompts for automated workflows
- Multi-reference editing — Character consistency, style transfer, and pose guidance
- Up to 4MP output — High-resolution generation at any aspect ratio
API Deprecations
October 31, 2025 - Flux Pro 1.0 Endpoints
The following endpoints will be deprecated. Migrate to flux-pro-1.1 or flux-kontext-pro.Deprecated endpoints:flux-pro-1.0flux-pro-1.0-depthflux-pro-1.0-canny
October 31, 2025 - Finetuning API
The finetuning API will be deprecated, including all finetuning-related endpoints.Deprecated endpoints:/v1/finetuneflux-pro-finetunedflux-pro-1.1-ultra-finetunedflux-pro-1.0-depth-finetunedflux-pro-1.0-canny-finetunedflux-pro-1.0-fill-finetuned

