Train a LoRA on your products and brand style with open tools on a local GPU

What a LoRA does for your images

A LoRA lets a company teach an open image model its own products or visual style, so that generated images show the real product and look like the brand. You collect 15 to 40 good photos, the range Black Forest Labs (BFL) gives for a style, describe each one in a short caption and train with a tool such as ai-toolkit; BFL's own example is a run of 1,800 steps. On FLUX.2 [klein] Base 4B or Qwen-Image, both under Apache 2.0, you may use the result commercially, and a workstation with one NVIDIA RTX PRO 6000 Blackwell Workstation Edition or Max-Q trains either model without memory-saving options.

LoRA stands for low-rank adaptation. The base model stays unchanged; training produces a small extra file of weights that the image program loads next to it, and that file can be switched on or off per job. An agency keeps one LoRA per client style, a manufacturer one per product line.

This guide is for marketing and brand managers and agency owners who decide whether to do this, and for the IT person or designer who will run the training. It covers the base models and their licences, the dataset, the tools, memory and time, and the hardware. Fine-tuning language models is a different subject, covered in our guide to GPU memory for LoRA and QLoRA. For using a finished LoRA on product shots, see our guide to AI product photos and packshots.

Base models whose licence allows commercial LoRAs

The licence of the base model decides what you may do with the LoRA and its images, so check it before the first photo is captioned.

Hugging Face cards and LICENSE files of Black Forest Labs and Qwen; training memory from Musubi Tuner's Qwen-Image documentation (1024 × 1024, batch size 1), BFL's LoRA guide of 4 June 2026 and SimpleTuner's FLUX.2 quickstart; all read on 30 October 2026.

FLUX.2 [klein] Base 4B is the undistilled version of BFL's small model, "Trained without step or guidance distillation". Its card calls it "Ideal for fine-tuning, LoRA training, research, and custom pipelines", says it "fits in ~13GB VRAM" for generation and is under Apache 2.0, and BFL's own LoRA guide states: "FLUX.2-klein-base-4B is Apache 2.0, so you can ship what you train."

Qwen-Image and its December update Qwen-Image-2512 are 20-billion-parameter models under Apache 2.0, as is the editing model Qwen-Image-Edit-2511, whose card lists "integrated LoRA capabilities". Qwen's card for the December update describes more natural detail and more accurate text rendering. The 20B models need more memory to train than the 4B klein model.

Several popular models are not for commercial LoRAs. The FLUX.2 [klein] 9B models are under the FLUX Non-Commercial License. FLUX.1 [dev] and FLUX.2 [dev] are under non-commercial licences too; BFL sells commercial licences for some of its models, as our FLUX GPU guide explains. Qwen-Image-2.1, released with a licence dated 20 September 2026, says "FOR NON-COMMERCIAL PURPOSES ONLY". FLUX 3 Image Klein Base, published in October 2026, allows "commercial output generation by Qualifying Users with less than US$5M in gross annualized revenue", affiliates included, and such users must apply content filters or review the outputs. Its licence grants the creation of fine-tuned versions, a LoRA included, only for non-commercial purposes, and its card says the exception is not blanket permission for commercial fine-tuning; larger companies need an arrangement with BFL. Whether a particular use is covered is for your legal counsel to confirm.

Photos, captions and image rights for the dataset

BFL's LoRA guide gives the clearest published numbers. A style LoRA needs "15–40 images that share one look", at least 1024 pixels on the long edge, each with a text file of the same name holding its caption. Captions describe what is in the image and say nothing about the style, and the guide advises a trigger word: "Pick a trigger word that is not a real word." The trigger word then calls up the learned look in a prompt.

A product LoRA follows the same logic, with the product as the fixed element. Photograph the one product from several angles and distances, on different backgrounds and in different light, and caption what changes (the setting, the angle, the light) while the trigger word stands for the product. None of the guides we read gives a separate image count for a single product, so start within BFL's 15 to 40 and judge the samples.

An edit LoRA learns a transformation from pairs of images. BFL's guide says "50–200 pairs is plenty", and its captions are instructions. A photo studio with an archive of raw shots and their retouched packshots already holds such pairs: an edit LoRA trained on them learns the house retouching style.

ai-toolkit reads images with a caption file of the same name, image2.jpg with image2.txt; a [trigger] placeholder in captions is replaced by the trigger word from the configuration, and the README says "You do not need to crop/resize your images."

Rights come first. BFL's guide names your own art, licensed photos and public-domain works as sources; train only on photos your company may use for this purpose: shots from an agency or a photographer may carry usage terms that do not cover AI training, and photos that show people raise data-protection questions. Your legal counsel and data-protection officer confirm. Whether fine-tuning can make a company the provider of a general-purpose AI model under the AI Act is explained, with the Commission's compute criterion, in our article on the AI Act and fine-tuning.

ai-toolkit, Musubi Tuner, SimpleTuner and kohya_ss compared

GitHub repositories and documentation of ostris/ai-toolkit, kohya-ss/musubi-tuner, bghira/SimpleTuner, kohya-ss/sd-scripts and bmaltais/kohya_ss, read on 30 October 2026. We give status and dates instead of version numbers.

ai-toolkit is the trainer BFL uses in its own guide. Its model list already includes FLUX 3 Image Klein from October 2026 as well as Qwen-Image, Qwen-Image-2512 and Qwen-Image-Edit-2511. Its web UI runs on port 8675 and starts, stops and monitors jobs; on a shared machine, set a password for the UI with the AI_TOOLKIT_AUTH variable.

Musubi Tuner comes from the kohya-ss project. It trains Qwen-Image and, with the --model_version option, Qwen-Image-Edit, Edit-2509 and Edit-2511, and it documents the memory needed for each memory-saving option. It expects the BF16 model files: its documentation states "fp8_scaled and fp8_e4m3fn versions cannot be used". Its latest release came in September 2026. Most of its code is under Apache 2.0, but its README says the code under the hunyuan_model directory "is modified from HunyuanVideo and follows their license", and the same applies to the HunyuanVideo 1.5 folder. Tencent's licence for HunyuanVideo states: "THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION". For that reason we do not install Musubi Tuner on the systems we deliver; ai-toolkit trains the same Qwen-Image models.

SimpleTuner covers many model families and had its latest release on 1 October 2026. Its Qwen Image guide notes "the current default is 2.1", the non-commercial model, so set the model flavour explicitly for client work. The older kohya_ss interface for kohya's sd-scripts trains Stable Diffusion and FLUX.1 models; it does not list FLUX.2 or Qwen-Image.

You can order the workstation alone, the setup of ai-toolkit and ComfyUI alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

GPU memory and training time per run

Musubi Tuner publishes the most detailed memory table we found, for a Qwen-Image LoRA at 1024 × 1024, batch size 1, with BF16 mixed precision and gradient checkpointing: 42 GB without further options, 30 GB with --fp8_base --fp8_scaled, 24 GB with 16 of the model's blocks swapped to system memory and 12 GB with 45 swapped. It recommends 64 GB of system memory with block swapping and notes that "Qwen-Image-Edit training requires additional memory for the control images". Block swapping moves parts of the model between the card and system memory during training, which saves memory and costs time.

SimpleTuner's Qwen guide calls 24 GB "the absolute minimum" and adds "40GB+ is strongly recommended for a smoother experience". For FLUX.2 [klein] its quickstart lists about 12 GB in BF16 for the 4B model with 32 GB or more of system memory. BFL writes that "A LoRA run on the 4B model fits in 24 GB of VRAM".

Time depends on the number of steps, the resolution and the card. BFL's guide says that an 1,800-step run takes under an hour on a 24 GB consumer card of the generation before Blackwell, and that for most style LoRAs "the visual peak is around step 750–1500", with edit LoRAs peaking "between 1000 and 1750 steps". Its advice is to "Watch the samples, not the loss", because the loss keeps falling after the images start to overfit. No training time is published for RTX PRO cards or for Qwen-Image LoRAs, so plan on a first test run on the delivered system.

Workstations and servers for LoRA training

We count usable memory by the rule in our guide on how many users one RTX PRO 6000 serves: 90 per cent of what the driver reports, less 3 GiB per card. The tools state GB; read as 10^9 bytes, 42 GB are 39.1 GiB by our arithmetic.

training memory as published by Musubi Tuner, SimpleTuner and BFL (see above); card memory from NVIDIA's datasheets; usable memory by our rule.

Entry. The 32 GB RTX PRO 4500 Blackwell (200 W) leaves 25.7 GiB by our rule, above the 24 GB that BFL gives for a LoRA on the 4B klein model. A Qwen-Image LoRA fits only with FP8 and block swapping, the 24 GB row of Musubi Tuner's table, which slows each step. The 24 GB RTX PRO 4000 Blackwell is the card size BFL names, but by our rule it leaves 18.5 GiB, so we start the ladder at the 32 GB card.

Entry, RTX PRO 5000. The RTX PRO 5000 Blackwell with 48 GB leaves 40.0 GiB and holds Musubi Tuner's 30 GB FP8 setting for Qwen-Image without block swapping. The 72 GB version leaves 61.5 GiB, enough for the full 42 GB run with about 22 GiB above it.

Mid. The 96 GB RTX PRO 6000 Blackwell leaves 83.0 GiB, about 44 GiB above the full Qwen-Image run, for the control images of an Edit-2511 LoRA, whose extra need Musubi Tuner does not quantify. A second card lets the team generate with yesterday's LoRA while today's trains. Two Workstation Edition cards at 600 W each are the realistic maximum for that edition; three or four Max-Q cards belong only in a chassis built for them with strong directed airflow. For system memory, Musubi Tuner and SimpleTuner recommend at least 64 GB for Qwen-Image; our workstation configurations follow NVIDIA's rule of at least twice the GPU memory, so 192 GB next to one 96 GB card and 384 GB next to two, as our guide to how much RAM a GPU server needs explains.

Large. An agency that trains a LoRA for each new client or campaign runs them as a queue on a server with RTX PRO 6000 Server Edition cards, one job per card. Four cards at 600 W with two AMD EPYC processors draw about 3.4 kW by our estimate and need a server room with a 32 A or three-phase feed.

Tell us which base model, how many LoRAs a month and how many people generate with them; we propose the workstation or server and the training setup that fits.

From the LoRA file to the image workflow

ComfyUI loads LoRAs next to the base model, and the same workstation that trains them can generate with them. BFL recommends training on the base checkpoint and generating with the distilled FLUX.2 [klein] 4B at 4 steps: "Applying the LoRA on the distilled model typically gives better results than the base model, and it's faster." Qwen-Image-Edit-2511 takes LoRAs directly, by its card. BFL's card for the 4B model encourages deployers to use the NSFW and protected-content filters in its repository, and its inference code includes an example of a watermark in the pixels and links the C2PA metadata standard, which helps with the machine-readable marking of Article 50(2).

Treat each LoRA as a production asset. Store the dataset, the captions, the configuration, the base model version and the chosen checkpoint together; the configuration in BFL's guide saves a checkpoint every 250 steps with sample images, and the best one is "not necessarily the last one". A new base model version means a new training run, so keep the dataset. Our article on backing up an AI server covers weights and adapters.

Generated product images are checked against the real product before they go online, as with any generated image; our product photo guide describes a composite that keeps the original product pixels. A realistic image of a real product can also meet the AI Act's definition of a deep fake, content that "resembles existing persons, objects, places, entities or events" and would falsely appear authentic, which deployers disclose under Article 50(4) from 2 August 2026; your legal counsel confirms what each campaign needs.

FAQ