AI product photos on your own GPU, from background removal to catalogue packshots

Product photos on your own workstation

AI product photography runs on a company's own workstation with open models: cut-outs on white, the product placed in a lifestyle scene, and a consistent packshot series for the catalogue. BiRefNet removes the background under the MIT licence, and FLUX.2 [klein] 4B and Qwen-Image-Edit-2511, both under Apache 2.0, build scenes and edit images for commercial use. A workstation with one NVIDIA RTX PRO 4000 Blackwell (24 GB) handles cut-outs and the small FLUX model; one RTX PRO 6000 Blackwell Workstation Edition or Max-Q (96 GB) holds the large editing model in full 16-bit precision next to the others.

This guide is for e-commerce and marketing managers, product photographers, agencies and the IT person who runs the machine. It covers the models and their licences, three workflows, how to keep the product unchanged, batch runs for a whole catalogue and the hardware.

A few terms first. A cut-out (German "Freisteller") is the product with its background removed, stored with a transparent background or a mask, a black-and-white image that marks which pixels belong to the product. A packshot is the clean, evenly lit product photo used on the product page. An image model generates or changes pictures from a written instruction, the prompt. ComfyUI is the free, open-source program in which these steps are connected into a workflow that a marketer can run again and again. BF16 and FP8 are number formats for a model's weights: BF16 uses 16 bits per value, FP8 uses 8 bits, which halves the files and can cost some image quality.

Our guide to the e-commerce AI server puts image editing next to product texts and search on one server, and our article on product videos from photos turns finished packshots into clips. This article covers the still images.

Models for cut-outs, scenes and edits

BiRefNet repository and model card, Hugging Face cards of Black Forest Labs and Qwen, and file sizes in Comfy-Org's Qwen-Image-Edit and Qwen-Image repackaging, all read on 30 October 2026. The Qwen-Image-Edit-2511 figure is the sum of the BF16 transformer (40.9 GB), text encoder (16.6 GB) and VAE (254 MB).

BiRefNet is a segmentation model with about 0.2 billion parameters: it sorts every pixel into product or background. Its repository lists 3.5 GB of GPU memory for inference in FP16 or BF16 and 4.8 GB in FP32, and a variant trained at 2048 × 2048 for high-resolution photos. ComfyUI's documentation states that "BiRefNet is natively supported in ComfyUI": the template "BiRefNet: Remove Background" returns the product on a transparent background together with its mask.

FLUX.2 [klein] 4B from Black Forest Labs (BFL) combines "text-to-image and image-to-image multi-reference editing in a single unified model". Its card says it "fits in ~13GB VRAM" and states "Open weights available for commercial use under the Apache 2.0 license". The example on the card uses 4 steps, a tenth of the 40 in the Qwen example below, so each image takes far fewer passes through the model.

Qwen-Image-Edit-2511 is Alibaba's 20-billion-parameter editing model. Its card lists the aims "mitigate image drift, improved character consistency", "integrated LoRA capabilities" and "enhanced industrial design generation". ComfyUI's tutorial notes that the 2509 version "adds multi-image input", so one picture can be edited with others as references, for example a product photo with a photo of a room. The card's example uses 40 steps.

Which models a shop may use commercially

Product photos are commercial use, so the licence comes before the hardware. BiRefNet is MIT, FLUX.2 [klein] 4B and Qwen-Image-Edit-2511 are Apache 2.0, and none of them adds a territory or revenue condition.

Several popular alternatives are restricted. BRIA's RMBG-2.0, built on the BiRefNet architecture and trained on "over 15,000" licensed images, is under CC BY-NC 4.0, and its card states: "Commercial use is subject to a commercial agreement with BRIA." This matters because rembg, a widely used MIT-licensed command-line tool for cut-outs, lists BRIA's model as "The default" and notes that RMBG-2.0 "requires a paid agreement for commercial use"; its README adds that "Model weights carry their own licenses, independent of rembg's MIT license". Pass -m birefnet-general and rembg uses BiRefNet instead.

FLUX.1 Kontext [dev] and the FLUX.2 [klein] 9B models are under non-commercial licences, FLUX 3 Image Klein allows commercial output only for companies below US$5 million of revenue, affiliates included, and Qwen-Image-2.1 is licensed "FOR NON-COMMERCIAL PURPOSES ONLY"; our FLUX GPU guide sets out BFL's terms and paid licences. None of these is part of the setup we deliver, and whether a specific use is covered is a question for your legal counsel.

Three ComfyUI workflows for product images

ComfyUI (GPL-3.0, actively developed, latest release 27 October 2026) lists Flux.2 Klein and Qwen Image Edit among its image editing models. Its App Mode exposes a finished workflow through a simple interface, so the marketer picks a photo and a scene and presses run.

1. Cut-out on white. Load the photo, run the BiRefNet template and place the result on white or on a transparent background. This is the classic "Freisteller" for marketplaces and catalogues, and the model does not redraw the product; it only decides which pixels belong to it.
2. Product in a scene. Generate the scene with FLUX.2 [klein] 4B, or give Qwen-Image-Edit-2511 the product photo and a room photo as two inputs with an instruction such as "place the kettle on the kitchen worktop, morning light from the left". The editing model redraws the product as part of the new picture.
3. Consistent series. For 200 products in the same look, keep one workflow, the same scene prompts, the same seed (the random start value) and the same output size, and change only the product photo. A LoRA trained on your products or house style, a small add-on file for the model, keeps the look closer still; our guide to training a LoRA on your brand and products covers that step.

You can order the workstation alone, the setup of ComfyUI with BiRefNet, FLUX.2 [klein] 4B and Qwen-Image-Edit 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.

Keeping the product exactly as it is

A generative model draws every pixel again. None of the model cards we read promises that a label, a logo or the shape of the product stays unchanged, and Qwen's own aim to "mitigate image drift" shows that the problem is known. A product photo with a different label or a wrong number of buttons no longer shows the product you deliver.

For product pages, a composite is the safer route. BiRefNet cuts out the product, the image model generates only the background or the scene, and ComfyUI's built-in node Image Composite Masked places the original product pixels on top, using "an optional mask that specifies which parts of the source image should be composited onto the destination image". The label stays exactly as photographed. A pasted product can look stuck on, because its shadow and the reflections on the new surface are missing; an editing pass can add them, but it redraws pixels again, so compare label and outline with the original before the image goes live.

Text in images needs the same care. ComfyUI's tutorial describes "precise bilingual text editing" for the Qwen-Image-Edit models, and the 2511 card's language tags are English and Chinese; it claims nothing for German, French or Polish, so test labels in your languages. For disclosure, the AI Act defines a deep fake as AI-generated or manipulated content that "resembles existing persons, objects, places, entities or events" and "would falsely appear to a person to be authentic or truthful"; Article 50(4), applicable from 2 August 2026, requires deployers to disclose such content, at the latest at first exposure. Whether a product scene falls under it is a question for your legal counsel; our article on the EU AI Act gives general information.

Catalogue runs for hundreds or thousands of products

A catalogue runs as a queue. ComfyUI describes "asynchronous queueing" and "a local API for integrating workflows into applications", so a script or the shop system submits one job per product. ComfyUI's BiRefNet template "processes a single image at a time"; for folders, rembg's rembg p command processes an input folder into an output folder, and its watch option -w picks up new files as the photographer drops them in.

The cut-out is the fast step. BiRefNet's repository gives 69.4 ms per image at 1024 × 1024 in FP16 on an A100, a data-centre card of an earlier generation, and calls its runtime figures a reference only; that is about 14 images per second by our arithmetic. No figure is published for RTX PRO cards.

The generative steps set the pace. For empty scenes generated with Qwen-Image at 1328 × 1328, our Qwen-Image server article estimates at most about 216 images per hour on one RTX PRO 6000 Workstation Edition with 50 steps, about 141 if the pipeline computes in BF16, and about 2,700 with the 8-step Lightning LoRA. All three are upper bounds by our estimate, derived from StorageReview's matrix-multiplication measurement on that edition (14 October 2026) and, for BF16, from NVIDIA's peak rate; real pipelines run slower. A shop seldom needs a new scene per product: by the same estimate, ten scenes reused for 2,000 composites take a few minutes of generation, while a workflow that edits every product photo costs one edit per product. For edits, which also process the input photos, no time is published for any RTX PRO card. Run 20 of your own products on the delivered system, measure the minutes, and multiply.

Send us the number of products, the image formats and how often the range changes; we size the workstation or server to that volume and name the models we would set up.

GPUs from one designer to a catalogue team

Memory decides which model and which precision fit on the card. 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 models need the file sizes from the table above, with working memory for the image on top.

memory from BiRefNet's repository and BFL's card; file sizes from Comfy-Org's repackaging on Hugging Face, read 30 October 2026; card memory from NVIDIA's datasheets; usable memory by our rule.

Entry. BiRefNet (3.5 GB) and FLUX.2 [klein] 4B (about 13 GB) come to about 16.5 GB, or 15.4 GiB, by our arithmetic and fit together within the 18.5 GiB that the 24 GB RTX PRO 4000 Blackwell (145 W, single slot) leaves by our rule. The 32 GB RTX PRO 4500 Blackwell leaves 25.7 GiB. That holds the FP8 transformer of Qwen-Image-Edit-2511 (20.5 GB), the part that draws the image, but not the whole FP8 set: with the text encoder (9.38 GB), which turns the prompt into numbers the model works with, and the VAE (254 MB), which turns the result into pixels, it comes to about 30.1 GB or 28.1 GiB. For regular edits, start at 48 GB.

Entry, RTX PRO 5000. The RTX PRO 5000 Blackwell with 48 GB leaves 40.0 GiB by our rule and holds the 28.1 GiB FP8 set with about 12 GiB to spare for the input images. Check FP8 quality on your own products before you decide on it. The 72 GB version leaves 61.5 GiB and holds the full BF16 files of Qwen-Image-Edit-2511, 57.8 GB or 53.8 GiB, with about 8 GiB to spare.

Mid. A 96 GB RTX PRO 6000 Blackwell leaves 83.0 GiB by our rule. BiRefNet, FLUX.2 [klein] 4B and the BF16 editing model stay loaded together, about 74 GB or 69 GiB by our arithmetic, so switching between workflows costs no loading time. A second card runs two jobs at once. 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.

Large. A server with RTX PRO 6000 Server Edition cards runs one queue per card for seasonal catalogue changes and can give a card to a language model for product texts. Four cards at 600 W with two AMD EPYC processors draw about 3.4 kW by our estimate, which needs a 32 A or three-phase feed in a server room. Our creative agency AI server shares such a server with video and rendering.

The software we set up, and your data

When we set up the software, the workstation gets ComfyUI, the BiRefNet template, FLUX.2 [klein] 4B and Qwen-Image-Edit-2511, and rembg with a BiRefNet model for folder runs. Each of these licences allows commercial use as described above; models with non-commercial or revenue-limited licences are not part of the setup.

Unreleased products, supplier photos and finished images stay on the workstation; whether a supplier's photos may be edited or composited depends on the rights the supplier grants, which your legal counsel confirms. The internet is needed only to download models and updates. Store the workflow, the prompt, the seed and the model version with every image, so an image can be produced again after a packaging change, and keep the original photographs: they are the reference against which every generated image is checked.

FAQ