FLUX 3 vs FLUX 2: Image Generation and Upgrade Comparison

Upgrading from FLUX.2 to FLUX 3 Image makes sense for teams that require native 4K generation, precise bounding-box layout editing, or self-hosting with FLUX 3 Image Klein. In Black Forest Labs (BFL) internal benchmark evaluations, FLUX 3 Image Klein scored higher than FLUX.2 [klein] 9B while running on fewer parameters. Billing also changes, from per-megapixel (MP) rates on FLUX.2 to fixed per-image prices by resolution class on FLUX 3.

The previous generation FLUX.2 launched in November 2025 across multiple variants including [pro], [max], [flex], and [klein]. FLUX.2 remains more economical for standard 1 MP to 2 MP images generated over an Application Programming Interface (API), charging $0.03 for the first megapixel on [pro]. Black Forest Labs introduced the FLUX 3 multimodal model family on July 23, 2026, followed by the release of FLUX 3 Image in early October 2026 and FLUX 3 Image Klein on October 15, 2026.

Three factors decide the switch: control features, local GPU memory, and monthly API volume. Pipelines that rely on immutable snapshot endpoints or the permissive Apache 2.0 license on FLUX.2 [klein] 4B should keep their existing configurations. Workflows that require sub-region image editing, multi-reference generation up to 10 images, or lower consumer Video Random-Access Memory (VRAM) footprints can transition to the FLUX 3 architecture.

FLUX 3 Image (and Klein) vs. FLUX.2: Side-by-Side

Are you one of these vendors? Update your listing

Short Upgrade Verdict by Production Use Case

Move to FLUX 3 if you need box-level edits or native 4K output. Stay on FLUX.2 if you want the lower per-megapixel API price on [pro] or fixed snapshot endpoints. In internal evaluations by Black Forest Labs, FLUX 3 Image Klein outperformed FLUX.2 [klein] 9B by 91 Elo points in text-to-image and 70 Elo points in editing despite using 38 percent fewer parameters. On the hosted Application Programming Interface (API), FLUX 3 Image Klein generates standard 1K images for $0.009, making it cheaper than FLUX.2 [klein] 4B at $0.014 for 1 megapixel.

For enterprise pipelines using [FLUX.2 [pro]](https://docs.bfl.ml/flux_2/flux2_overview), the cost calculation differs. FLUX.2 [pro] charges $0.03 for the first megapixel and $0.015 for each additional megapixel, whereas FLUX 3 Image costs $0.048 for a 1K image and $0.10 for a 2K image. Teams that do not require bounding-box control or native 4K canvas sizes keep lower operational costs on FLUX.2 [pro].

Inside FLUX.2's 4 MP ceiling there is no price break-even in favor of FLUX 3 Image. FLUX.2 [pro] costs $0.03 at 1 MP, $0.045 at 2 MP and $0.075 at 4 MP, against $0.048, $0.07 and $0.10 for the 1K, 1.5K and 2K classes on FLUX 3 Image. On the small tier FLUX 3 Image Klein is cheaper, at $0.009, $0.011 and $0.014 for those sizes against $0.014, $0.015 and $0.017 on FLUX.2 [klein] 4B.

• Self-hosters running FLUX.2 [klein] 9B or 4B: Move to FLUX 3 Image Klein for BFL-reported gains over [klein] 9B and an 8 GB INT8 generator (the local prompt upsampler needs about 28 GB more), if your organization earns under $5 million a year or buys a self-hosted license. [klein] 4B users who need Apache 2.0 terms with no revenue cap should stay on FLUX.2.
• Hosted API teams on FLUX.2 [pro]: Remain on FLUX.2 [pro] for high-volume 1 MP to 4 MP generation if your pipeline relies on the lower $0.03 base rate and does not need bounding-box edits or native 4K.
• Teams requiring 4K resolution or bounding-box edits: Upgrade to FLUX 3 Image or FLUX 3 Image Klein for coordinate-driven placement on a 0 to 1000 grid and native output up to 16.8 megapixels.
• Workflows requiring reproducible snapshot outputs: Remain on FLUX.2 to keep using fixed snapshot endpoints like flux-2-pro and flux-2-klein-9b that Black Forest Labs keeps for reproducible workflows.
• Teams planning video production: Consider FLUX 3 Video, which is generally available on BFL's API, for clips up to 20 seconds with synchronized audio, an asset type that FLUX.2 does not generate.

Architectural Changes from FLUX.2 to FLUX 3 Image

FLUX 3 Image replaces the 4-megapixel ceiling of FLUX.2 with native 4K generation, expands reference inputs to 10 images, and introduces coordinate bounding boxes for non-destructive editing. While FLUX.2 capped maximum canvas areas at 4 megapixels across its [pro], [max], and [klein] variants, FLUX 3 Image generates native 2K and 4K images directly. In documentation from Black Forest Labs, the 4K mode reaches 5456 by 3072 pixels, representing 16.8 megapixels of rendered detail without external upscaling.

FLUX 3 Image and FLUX 3 Image Klein drop FLUX.2's tier-by-tier reference limits. Under FLUX.2, reference allowances varied by model tier: [klein] accepted up to 4 images, [dev] accepted up to 6, and [pro], [max], and [flex] accepted up to 8 via the API (10 in the playground). In FLUX 3 Image and FLUX 3 Image Klein, every tier accepts up to 10 reference images, tagged directly within prompt text using structured references from ref_image_0 through ref_image_9.

Bounding boxes are a control FLUX 3 Image adds that BFL's FLUX.2 docs do not list. Black Forest Labs introduced bounding-box coordinate controls that operate on a normalized 0 to 1000 grid on both axes, documented in the FLUX 3 Image bounding boxes guide. Developers pass coordinates to place, recolor, move, replace, or remove specific visual elements while preserving the surrounding composition, a workflow supported by an integrated prompt upsampler that elaborates descriptions without altering spatial box boundaries.

• Canvas expansion: FLUX 3 delivers native 4K output up to 5456 by 3072 pixels (16.8 MP), compared to the 4 MP ceiling on FLUX.2.
• Uniform reference handling: FLUX 3 Image and FLUX 3 Image Klein both accept up to 10 reference images, removing the tier-based limits found in FLUX.2.
• Spatial layout control: Bounding boxes specified on a 0 to 1000 grid enable localized modifications while leaving adjacent pixels unchanged.
• Consolidated endpoints: A single API endpoint handles standard text-to-image, single-image editing, and multi-reference composition.

Klein Model Comparison and Hardware Requirements

FLUX 3 Image Klein is a single 5.6-billion-parameter open model whose INT8 build runs on 8 GB of VRAM, and in BFL's internal tests it rates above FLUX.2 [klein] 9B. FLUX.2 separated its lightweight tier into two distinct parameter sizes: a 4-billion-parameter model running under Apache 2.0 and a 9-billion-parameter model under the FLUX Non-Commercial License. Both FLUX.2 [klein] models are 4-step distilled networks. BFL lists the 4B at about 13 GB of VRAM on consumer GPUs.

By contrast, FLUX 3 Image Klein features 5.6 billion parameters in its image generation backbone and ships in two formats on Hugging Face: Klein (a 16-step distilled model running at guidance 1 for rapid generation) and Klein Base (a 50-step undistilled model running at guidance 3.5 for research and fine-tuning). In BFL internal evaluations, Klein ranked first among evaluated open models in text-to-image, multi-reference generation, and editing, scoring higher than proprietary systems including Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B, Klein achieved a 91 Elo point gain in text-to-image and a 70 Elo point gain in editing while reducing parameter count by 38 percent.

Local hardware accessibility has improved through quantized builds created in partnership with NVIDIA, though prompt expansion introduces a hardware trade-off. An INT8 quantized build allows Klein to run on NVIDIA GeForce RTX Graphics Processing Units (GPUs) with as little as 8 GB of VRAM using ComfyUI or Diffusers. However, Black Forest Labs warns that Klein requires its companion prompt upsampler to produce optimal visual quality. Because the local upsampler model adds roughly 28 GB of memory consumption, running the complete pipeline entirely on a local machine requires far more memory unless you write full scene captions yourself.

Pricing Comparison and Resolution Economics

FLUX 3 Image Klein undercuts FLUX.2 [klein] 4B at 1, 2 and 4 MP, while FLUX 3 Image costs more than FLUX.2 [pro] at every size both produce. Under the published FLUX.2 overview, API costs scale by megapixel area. FLUX.2 [pro] charges $0.03 for the first megapixel and $0.015 for each additional megapixel, while FLUX.2 [klein] 4B charges $0.014 for the first megapixel and $0.001 per additional megapixel.

As outlined on the BFL API pricing documentation, FLUX 3 Image charges fixed prices across resolution classes: 768 by 768 pixels costs $0.041, 1K (about 1 MP) costs $0.048, 1.5K (about 2 MP) costs $0.07, 2K (about 4 MP) costs $0.10, and native 4K (about 16.8 MP) costs $0.607. Conversely, FLUX 3 Image Klein API generation costs $0.008 for 768 by 768 pixels, $0.009 for 1K, $0.011 for 1.5K, $0.014 for 2K, and $0.036 for 4K. Both text-to-image and image editing requests share identical pricing on the FLUX 3 endpoints.

At standard 1 MP resolution, generating 1,000 images costs $30 on FLUX.2 [pro], $48 on FLUX 3 Image, and $9 on FLUX 3 Image Klein. At 4 MP resolution, FLUX.2 [pro] costs $0.075 per image ($0.03 for the base megapixel plus $0.045 for three additional megapixels), making it cheaper than FLUX 3 Image at $0.10 for 2K output. At 4K resolution, FLUX 3 Image Klein generates 1,000 images for $36, while FLUX 3 Image costs $607; FLUX.2 does not offer a native 4K tier.

• 1 MP generation: FLUX 3 Image costs $0.048, FLUX.2 [pro] costs $0.030, and FLUX 3 Image Klein costs $0.009.
• 4 MP generation: FLUX.2 [pro] costs $0.075, FLUX 3 Image 2K costs $0.100, and FLUX 3 Image Klein 2K costs $0.014.
• Native 4K generation: FLUX 3 Image costs $0.607 per image, FLUX 3 Image Klein costs $0.036 per image, and FLUX.2 has no equivalent tier.
• No free tier: BFL's API pricing lists no free credits, so the only route without per-image fees is self-hosting Klein under the FLUX Kommunity License on your own GPU.

→ Compare per-token pricing for every AI model

Capabilities and Features Retained in FLUX.2

Upgrading entirely to FLUX 3 requires abandoning Apache 2.0 open-source licensing, fixed snapshot API endpoints, and specialized FLUX.2 grounding features. Black Forest Labs continues to host FLUX.2 endpoints, and its FLUX.2 docs still recommend FLUX.2 for many text-to-image tasks. Pipelines that prioritize long-term output consistency benefit from immutable snapshot endpoints, such as flux-2-pro and flux-2-klein-9b, which protect automated systems from unintended shifts in model weights.

The license gap matters most to companies above $5 million in revenue. FLUX.2 [klein] 4B was released under the permissive Apache 2.0 license, allowing unrestricted commercial use, modification, and redistribution regardless of company size. In contrast, FLUX 3 Image Klein is governed by the FLUX Kommunity License, which restricts free commercial deployment to entities with annual revenue below $5 million. Organizations exceeding that threshold must negotiate commercial licenses with Black Forest Labs sales or pay per-image API rates.

Specific features available in FLUX.2 variants remain absent from the FLUX 3 documentation. FLUX.2 [max] offers grounding search, which BFL's FLUX 3 Image pages do not list. FLUX.2 ships open weights for [klein] 4B and 9B, while FLUX 3 Image has no public weights, and the multimodal FLUX 3 Dev backbone has been announced without a release date.

• Snapshot reproducibility: FLUX.2 maintains immutable snapshot endpoints for reproducible generation pipelines.
• Uncapped open-source licensing: FLUX.2 [klein] 4B provides an unrestricted Apache 2.0 license without revenue limits.
• Grounding search: FLUX.2 [max] offers grounding search, which BFL's FLUX 3 Image pages do not list.
• Weight availability: FLUX 3 Image offers no public open weights, requiring commercial licensing for self-hosted execution.

Pipeline Migration Checklist

Moving a pipeline from FLUX.2 to FLUX 3 means updating API endpoints, converting megapixel dimensions into resolution classes, and re-testing any custom Low-Rank Adaptation (LoRA) weights. Hosted API calls move to POST api.bfl.ai/v1/flux-3-image or POST api.bfl.ai/v1/flux-3-image-klein, and each endpoint handles both text-to-image and editing.

Payload formatting requires structural adjustments to handle resolution classes and image constraints. Developers must replace custom width and height megapixel calculations with discrete resolution tiers (768sq, 1k, 1.5k, 2k, 4k) paired with standard aspect ratios such as 16:9, 4:3, or 1:1. Pipelines incorporating reference images must enforce API constraints: inputs below 256 by 256 pixels trigger an HTTP 422 error, while reference files exceeding 16 megapixels return an HTTP 400 error.

Local deployment teams must address prompt expansion and weight compatibility before putting Klein into production. Running Klein via ComfyUI requires updating the software installation, downloading the Klein weights from the official repository, and importing the official ComfyUI workflow. BFL has not published whether FLUX.2 LoRA weights load on Klein. Test existing fine-tunes before production, and plan to retrain on FLUX 3 Image Klein Base, which BFL offers for LoRA training with day-one support in NVIDIA NeMo AutoModel.

• Update API endpoints: Point generation calls to api.bfl.ai/v1/flux-3-image or api.bfl.ai/v1/flux-3-image-klein.
• Map resolution classes: Convert continuous pixel targets to standard tiers (768sq, 1k, 1.5k, 2k, 4k) and predefined aspect ratios.
• Validate reference inputs: Ensure input images meet the minimum 256 by 256 pixel and maximum 16-megapixel validation boundaries.
• Re-test LoRA weights: BFL has not published whether FLUX.2 LoRAs load on Klein, so plan to train new adapters on the 50-step Klein Base model.
• Configure prompt expansion: Implement caption preprocessing or allocate sufficient memory for the 28 GB local prompt upsampler model.

The Verdict

FLUX 3 Image and FLUX 3 Image Klein represent clear technical upgrades for teams building interactive editing tools, multi-reference generation workflows, or high-resolution visual pipelines. The introduction of 0 to 1000 coordinate bounding boxes, native 4K canvas generation up to 16.8 megapixels, and up to 10 reference image inputs gives FLUX 3 controls that BFL's FLUX.2 docs do not list for API use. Self-hosted operators also gain a smaller model: the INT8 build of FLUX 3 Image Klein runs its image generator on 8 GB VRAM GPUs, and BFL's internal tests rate it above FLUX.2 [klein] 9B.

FLUX.2 [pro] costs less than FLUX 3 Image at every shared size. FLUX 3 Image Klein undercuts both at $0.009 per 1K image, and BFL's internal tests rate it above FLUX.2 [pro], so the price case for FLUX.2 [pro] holds only against the flagship. FLUX.2 [pro] generates 1-megapixel images for $0.03, undercutting FLUX 3 Image at $0.048, and scales to 4 megapixels for $0.075 compared to $0.10 for FLUX 3 Image 2K. Developers needing reproducible outputs also benefit from retaining FLUX.2, as Black Forest Labs maintains immutable snapshot endpoints that protect production systems from weight revisions.

Neither model tier fits organizations that demand fully unrestricted open-source terms alongside flagship model weights. FLUX 3 Image provides no public weights, restricting local hosting to commercial licensees, while FLUX 3 Image Klein restricts free commercial use to entities earning under $5 million annually. Teams requiring completely open enterprise licensing must remain on FLUX.2 [klein] 4B under Apache 2.0 or await future open releases.

A cut in FLUX 3 Image API prices, or an Apache-licensed FLUX 3 Dev release, would remove the price and license reasons to stay on FLUX.2, leaving fixed snapshot endpoints as the main one. Until then, volume and resolution decide the cost. At 100,000 images a month at 1 MP, FLUX.2 [pro] costs $3,000, FLUX 3 Image $4,800 and FLUX 3 Image Klein $900.