# FLUX Pro Finetuning API Guide: Retirement and Alternatives > Published 2026-04-06, updated 2026-09-05 · https://www.promptzone.com/kofi_choi/flux-pros-fine-tune-api-breakthrough-31l9 --- title: "FLUX Pro Finetuning API Guide: Retirement and Alternatives" published: true description: "Understand the FLUX Pro Finetuning API retirement and choose a documented FLUX.2 training or reference-editing workflow." tags: ai,imagegeneration,flux,lora --- The FLUX Pro Finetuning API was Black Forest Labs’ hosted service for adapting Pro image models to user-supplied concepts; BFL deprecated it on October 31, 2025, with no migration path. For new work, fal documents FLUX.2 dev LoRA training, while BFL documents Klein training and a separate Klein LoRA inference service. [Release notes][releases], [Trainer API][trainer], [Klein training][klein], [Klein LoRA inference][inference] ## What are the key facts about the FLUX Pro Finetuning API? | Field | Verified detail | | --- | --- | | Developer | Black Forest Labs. [Launch][launch] | | Released | January 16, 2025; deprecation took effect October 31, 2025. [Launch][launch], [Release notes][releases] | | Type | Hosted customization and inference for FLUX Pro image models, now deprecated. [Launch][launch], [Release notes][releases] | | Size or parameters | Parameter count specific to the fine-tuning service: not published in its launch announcement. [Launch][launch] | | License and access | Hosted Pro service; no open Pro weights. BFL announced discontinuation with no migration path. [Model access][repo], [Release notes][releases] | | Where it runs | The original service ran on provider infrastructure through API endpoints; it was not a local Pro download. [Launch][launch], [Model access][repo] | ## Which customization options replace a FLUX Pro training workflow? The original service addressed a specific task: teaching a hosted image model concepts represented in a user’s example images. BFL demonstrated personalized objects, people, pets, clothing, and visual styles. [Launch][launch] Its launch announcement described applying the resulting customization to several Pro generation and editing tools. These were capabilities of the service at launch, not a statement that new training remains available today. [Launch][launch] The underlying creative need remains useful to define. Write down whether you need a recurring character, a repeatable visual style, a particular product, or a transformation between images before choosing another system. fal’s FLUX.2 training documentation distinguishes text-to-image adaptation from image-to-image training. The former teaches concepts or styles; the latter trains transformations from examples. [Training guide][falguide] BFL also documents training Klein Base models with open weights. That is an alternative for users who want a local customization workflow and can meet its setup requirements. [Klein training][klein] ## What did the FLUX Pro Finetuning API retirement affect? BFL’s October 2025 notice covers the FLUX.1-era training and fine-tuned generation endpoints, with no migration path. Its April 2026 release notes separately announce Klein LoRA inference; the older retirement does not describe that newer service. [Release notes][releases] The notice specifically covers `/v1/finetune` and several `*-finetuned` endpoints. Do not treat an old request example as a supported route to creating a new Pro fine-tune. [Release notes][releases] There is a documentation discrepancy: BFL’s current OpenAPI schema still contains some fine-tune utility and inference definitions. The dated deprecation notice remains essential context for interpreting them. [OpenAPI schema][schema], [Release notes][releases] The presence of a schema entry does not establish that a retired workflow is available to your account. For a new build, check the documentation for the specific model and endpoint. Plan a new training run from your retained source images when moving away from the retired Pro service; BFL announced no migration path for those fine-tunes. [Release notes][releases] For current alternatives, check the chosen model’s license and provider terms. Klein’s 4B and 9B Base models have different weight licenses, and a hosted trainer is a separate service. [Klein training][klein], [Trainer API][trainer] ## How do you train a new FLUX.2 LoRA through fal? The original Pro fine-tuning service is deprecated. The following is a concrete alternative using fal’s current FLUX.2 dev text-to-image LoRA trainer; it creates a new adaptation for that model. [Trainer API][trainer] First, decide what the dataset should teach. Gather examples that share the intended subject or style, then prepare the archive according to the trainer’s input schema. [Trainer API][trainer] The schema accepts a ZIP archive URL in `image_data_url`. Caption files can accompany images; when captions are absent, supply `default_caption`, because the documented trainer otherwise reports an error. [Trainer specification][spec] Create a fal API key and configure `FAL_KEY` in your server environment. Install `@fal-ai/client`, then supply your own archive URL through `TRAINING_ZIP_URL`. [Trainer API][trainer] ```javascript import { fal } from "@fal-ai/client"; const result = await fal.subscribe("fal-ai/flux-2-trainer", { input: { image_data_url: process.env.TRAINING_ZIP_URL, default_caption: "An illustration in the house drawing style", steps: 1000, output_lora_format: "comfy", }, }); console.log(result.data.diffusers_lora_file.url); ``` Replace the example caption with one appropriate to your dataset. The code invokes a paid training job when executed; review the trainer’s displayed price before submitting it. [Trainer specification][spec] The endpoint documents a charge of $0.0064 per training step, or $6.40 for the 1,000-step example. This is the current alternative’s training price, not pricing for the retired Pro service. [Trainer specification][spec] Download the returned LoRA and configuration artifacts. The provider’s training guide explains how to use the resulting adaptation with a compatible FLUX.2 inference workflow. [Trainer API][trainer], [Training guide][falguide] Before scaling a training run, reserve representative evaluation prompts. Check whether the result preserves the desired concept across new compositions rather than only reproducing the training images’ setting. If your task is a one-off change to an existing image, evaluate reference editing first. FLUX.2 dev supports character, object, and style references without additional fine-tuning. [Model card][dev] The sibling [FLUX.2 download guide](/thandi_fischer/flux2-ai-powerful-image-generation-model-unveiled-1mib) covers local access. The [LoRA pillar](/tara_suzuki/how-to-use-loras-in-comfyui-in-2026-load-stack-and-troubleshoot-235e) gives background for loading adapters. For a local training route, follow BFL’s Klein Base guide. BFL also documents uploading a trained Klein `.safetensors` LoRA through Dashboard → Customization → Finetunes, then using the matching hosted endpoint. This inference service is in public beta and requires a LoRA trained for the selected Klein base model. [Klein training][klein], [Klein LoRA inference][inference] ## How do current FLUX customization options compare? | Approach | What you are choosing | | --- | --- | | Original Pro Finetuning API | A deprecated hosted service with no announced migration path. [Release notes][releases] | | FLUX.2 dev trainer on fal | A separate hosted service that returns LoRA artifacts. [Trainer API][trainer] | | Klein training and BFL inference | Train a Klein LoRA, then run a compatible local workflow or upload it to BFL’s separate hosted inference beta. [Klein training][klein], [Klein LoRA inference][inference] | Compare them by present availability and deployment needs. The sibling [Kontext overview](/riya_ahmadi/flux-kontext-ai-model-debuts-3988) also explains a reference-editing approach to consider before training. ## What else should you know about FLUX Pro fine-tune retirement? ### Can I start a new FLUX Pro fine-tune through the original API? BFL deprecated the original FLUX Pro Finetuning API on October 31, 2025. For new customization, use a separately documented service such as fal’s FLUX.2 dev LoRA trainer. [Release notes][releases], [Trainer API][trainer] ### Were FLUX Pro fine-tunes downloadable open weights? FLUX Pro fine-tuning was a hosted service, and BFL does not list open Pro weights. Its fine-tune identifiers were used with the service’s generation endpoints. [Launch][launch], [Model access][repo] ### Can a FLUX Pro fine-tune migrate to a FLUX.2 trainer? BFL announced no migration path for the retired FLUX Pro Finetuning API. fal’s FLUX.2 trainer creates a new LoRA from a supplied dataset. [Release notes][releases], [Trainer API][trainer] ### Does BFL offer a current API for trained Klein LoRAs? BFL documents a separate FLUX.2 Klein LoRA inference service in public beta. Train a compatible Klein LoRA, upload it through the Dashboard, and call the matching fine-tuned endpoint with its identifier. [Klein LoRA inference][inference] ## Sources - [Original Pro Finetuning API launch][launch] - [BFL dated API deprecation notices][releases] - [FLUX.1 model and hosted access overview][repo] - [BFL current OpenAPI schema][schema] - [fal FLUX.2 training guide][falguide] - [fal FLUX.2 trainer API documentation][trainer] - [fal FLUX.2 trainer schema and pricing][spec] - [BFL Klein Base training guide][klein] - [FLUX.2 dev model card][dev] - [BFL Klein LoRA inference and Dashboard upload guide][inference] [launch]: https://bfl.ai/blog/25-01-16-finetuning [releases]: https://docs.bfl.ai/release-notes [repo]: https://github.com/black-forest-labs/flux [schema]: https://api.bfl.ai/openapi.json [falguide]: https://blog.fal.ai/training-flux-2-loras/ [trainer]: https://fal.ai/models/fal-ai/flux-2-trainer/api [spec]: https://fal.ai/models/fal-ai/flux-2-trainer/llms.txt [klein]: https://docs.bfl.ai/flux_2/flux2_klein_training [dev]: https://github.com/black-forest-labs/flux2/blob/main/model_cards/FLUX.2-dev.md [inference]: https://docs.bfl.ai/flux_2/flux2_lora_inference ## Related guides on PromptZone - [Realistic Photos with FLUX](/stabletom/realistic-photos-with-flux-57aa) - [Best SDXL Models in 2026](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116)