Crashes, No Sound, or Screen Glitches?
Free tools Windows power users keep installed
One-click scans. No signup required.
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall
ThatPainter is reader-supported. When you buy through links on our site, we may earn an affiliate commission. Learn More
Black Forest Labs really did launch a FLUX fine-tuning service that accepted as few as five images—but that was a January 2025 announcement, not a feature you can use through its original hosted API today. BFL says the Finetuning API was discontinued on October 31, 2025, with no migration path. The five-image figure was a minimum, not a guarantee of a reliable likeness, product replica, or style model.
What BFL announced in 2025
On January 16, 2025, Black Forest Labs announced the FLUX Pro Finetuning API, a hosted service for customizing FLUX from five to 20 training images. The launch described four training modes: character, product, style, and general. Users could optionally supply text descriptions, configure settings such as training iterations and learning rate, and use a trigger word to call up the learned concept in later prompts. VentureBeat’s launch coverage reported the announcement’s capabilities and requirements.
As an Amazon Associate I earn from qualifying purchases.
This was not a way to train a new image-generation foundation model from scratch. FLUX already supplied broad image-generation ability; fine-tuning aimed to associate that capability with a narrower concept, such as a particular product, character, or visual treatment. In practical terms, a base model might know what a sneaker is, while fine-tuning tries to teach it what your sneaker looks like in new scenes.
What “five images” meant—and what it didn’t
Five was the advertised lower bound for the 2025 service, not a universal recommendation or a promise that five arbitrary pictures would produce a dependable custom model. The smaller the dataset, the more each image can influence what the model learns. If all five show the same pose, crop, background, or lighting, the model may associate those incidental details with the intended subject.
#1 Best Overall
A small, carefully chosen set could plausibly help with a visually distinctive object, a character with a narrow design, a likeness for controlled creative experiments, or a tightly defined visual style. But five examples cannot show every angle, material, expression, outfit, lighting condition, or product configuration. The likely outcome was improved consistency in a narrow range of prompts—not perfect identity preservation or production-ready results without iteration.
BFL’s current guidance for training a style model on FLUX.2 Klein is more demanding: it recommends 20–40 images, says fewer than 20 may lack enough variation, and warns that a highly varied set above 40 can dilute the style. Its example uses 27 images. That is a different, current training workflow, not a revised minimum for the discontinued hosted API. See the FLUX.2 Klein training example.
How the launch-era workflow worked
The broad process was to upload a small image set, optionally add captions or descriptions, choose training settings and a trigger word, then wait for a fine-tuning job. Once trained, a user could supply the resulting fine-tune identifier to supported generation endpoints, put the trigger word in a prompt, and adjust how strongly the customization affected the result.
Launch-era reporting listed JPG, JPEG, PNG, and WebP support, with training images capped at about one megapixel for optimal results. The exact original beta commands should not be treated as a current how-to: BFL discontinued the relevant hosted endpoints. Historical API documentation shows controls including finetune_id and finetune_strength; one FLUX 1.1 Pro Ultra endpoint documented a strength range of 0 to 2 and a default of 1.2. These are historical interface details, not confirmation that those endpoints still accept requests. The endpoint reference preserves some of that context.
Choosing training images for a narrow concept
If you are evaluating a comparable custom-training workflow today, treat the dataset as a set of examples of what should remain consistent, not simply as a folder of attractive pictures. Use sharp, clean images with varied angles and distances, while keeping the defining subject traits visible. Avoid near-duplicates, heavy filters, watermarks, and distracting backgrounds. Where captions are supported, describe the subject and relevant visual attributes without accidentally tying the concept to one repeated scene.
- People: Get the person’s consent and consider privacy, publicity rights, platform rules, and how the resulting model could be used. Do not treat fine-tuning as a safe route to an “accurate digital twin.”
- Products: Include useful views, but do not assume a few photos reveal hidden sides, dimensions, materials, or internal details. Logos, small text, colors, and packaging can drift in generated images; review advertising and e-commerce imagery against the real product.
- Characters: Show the defining features from more than one pose or view if you expect them to persist in new scenes. A handful of similar portraits may teach a pose or framing as readily as a character identity.
- Styles: Include enough variation to teach a visual treatment rather than one composition. Consider rights to the source images and avoid assuming that training on a few works makes reproducing a living artist’s recognizable style legally or ethically unproblematic.
Trigger words also need testing. A common word can collide with ordinary prompt language; an overly descriptive one can tie the concept to unwanted attributes. At high influence, a model may over-apply the subject or style; too little influence may fail to retain it. Neither a trigger word nor a strength control removes the need to test varied prompts and inspect results.
What it was meant to enable
The 2025 launch positioned customization for brand imagery, product visualization, marketing, character art, and storytelling. Launch coverage also described use with image-generation and editing workflows, including FLUX.1 Fill for inpainting and FLUX.1 Depth for structural control, with supported workflows reaching up to four megapixels. Those are claims about the launch-era service and associated endpoints, not a statement about current availability.
The same distinction applies to price. Launch-era coverage listed generation prices of $0.06 per image for FLUX 1.1 [pro] Ultra, $0.04 for FLUX 1.1 [pro], $0.05 for FLUX.1 [pro], and $0.025 for FLUX.1 [dev]. Those were image-generation prices, not proof of the total cost of a fine-tuning job; training, storage, API credits, or other charges should not be inferred from them.
The hosted fine-tuning API is gone
BFL’s official release notes state that its Finetuning API and related endpoints were deprecated and that the functionality was discontinued on October 31, 2025. BFL also says there is no migration path. So a reader cannot sign up today and upload five images to recreate that original hosted workflow.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Some old endpoint documentation or references may remain visible, but documentation alone is not evidence that a discontinued service is operational. The newer FLUX.2 family is not a direct replacement for the hosted fine-tuning API.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What you can consider instead
BFL’s current documentation describes FLUX.2 Klein base variants as options for fine-tuning, LoRA training, research, and custom pipelines. That points to a more technical route rather than the old upload-and-train hosted service. The 4B version is described as Apache 2.0 and suited to consumer GPUs with roughly 13 GB of VRAM; the 9B version uses BFL’s FLUX NCL license. Check the applicable terms and current requirements in the FLUX.2 overview before building a commercial workflow.
For occasional use, multi-reference prompting with a current FLUX.2 model may avoid training a persistent adapter, though consistency across a large batch can be less predictable. For guided edits rather than a reusable custom model, BFL’s FLUX.1 Kontext overview describes image-and-text editing workflows. Its [dev] model is identified as non-commercial unless separately licensed. These options are not interchangeable: choose based on whether you need repeated subject consistency, image-guided edits, or just a one-off generation.
Current generation prices are separate from training. As listed in BFL’s pricing documentation observed in August 2026, one credit equals $0.01; FLUX.2 Klein starts around $0.014 per image for 4B and $0.015 for 9B, while Pro, Max, and Flex are priced per megapixel. Prices can change, and none of these figures revives or prices the discontinued fine-tuning service. Consult BFL’s current pricing page.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Three rights questions to settle
Before training or publishing generated work, separate three issues: whether you have rights and consent to upload the source images; what rights apply to the trained model or adapter; and what license and other restrictions apply to generated outputs and commercial deployment. API access does not by itself settle all three. FLUX model licenses differ, so check the exact model’s current terms rather than assuming that rights carry over from one version to another.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
