The test for AI product photography is not whether the image looks good. It is whether the label is still readable, the logo is still the right logo, and the bottle is still the shape your factory makes.
Most generation fails that test in a specific way: it produces something that looks like your product rather than being your product. Black Forest Labs' documented approach treats that as a reference problem, not a prompting one.
Sourcing note: the capability claims, workflow and prompt techniques below come from BFL's product consistency, product mockups and multi-reference editing documentation, plus their pricing rules. Last verified July 30, 2026.
What this guide solves
The pain point: "describe the product in the prompt" does not survive contact with real branding. No amount of text description reproduces a specific label layout or a particular bottle silhouette.
The differentiator: the reference-driven method BFL actually documents, the reference-budget maths that decides how many product angles you can supply, and the point at which the right answer is a Tool rather than a generation.
What FLUX preserves
BFL's product consistency documentation states it plainly: FLUX.2 keeps product identity intact while changing context, background or presentation. You upload a product photo or logo and describe the new setting, and the model preserves branding, labels, shapes and materials.
Their documented examples map neatly onto real e-commerce jobs:
- Logo on a new product — brand mark applied to a different item
- Product in a new scene — same item, different environment
- Product in unusual settings — including physically demanding ones like underwater
- Product on retail packaging — item rendered onto pack designs
- Logo on merchandise — brand applied across a merch range
The common shape: the identity comes from an image, the context comes from the prompt. Reverse that and you get generic product-looking output.
The workflow
1. Start from a clean reference
The reference does the identity work, so its quality sets your ceiling. A packshot on plain background, evenly lit, with the label fully legible, beats a lifestyle photo where the product is half in shadow.
If your only source imagery is cluttered, clean it first — this is exactly what FLUX Erase is for, removing stands, clamps and props with a mask and no prompt.
2. Budget your references
Multi-reference has a hard constraint: [pro] has a 9MP total budget across input and output. At 1MP output you get up to 8 references; at 2MP, 7. klein caps at 4.
For product work, that budget usually buys you: the product front, the product at an angle, the label close-up, and perhaps a material detail. That is a strong identity set. Spend the rest of the budget on output resolution only if you genuinely need it — generating at 1MP and upscaling protects consistency.
3. Assign roles in the prompt
BFL's multi-reference rule applies here more than anywhere: describe the role of each image so the model knows what to pull from where.
- "The bottle from image 1" — the product
- "on the marble surface of image 2" — the setting
- "with the label design from image 3" — the branding detail
Unassigned references get averaged, which is precisely how labels turn into label-shaped smudges.
4. Describe the photography, not just the scene
BFL's product mockup guidance is to describe the product, surface, lighting setup and composition. Their prompting guide is explicit that naming optics beats generic quality words — "shot on Hasselblad X2D, 80mm lens, f/2.8, natural lighting" outperforms "professional product photo."
For e-commerce specifically, that means specifying: surface material, light direction and softness, whether there are reflections, and how much of the frame the product occupies.
Need the product frames themselves? Flux 3 AI is a browser workspace for product visuals, packaging concepts and campaign frames, plus upscaling for delivery. Open the image generator or see the credit plans.
Getting text on packaging right
Labels are typography, and FLUX has documented techniques for it:
- Put the exact string in quotation marks — "the text 'ACME' appears on the front label"
- Specify placement relative to other elements
- Describe the style — "clean sans-serif", "embossed serif"
- Set the size — "large headline text", "small ingredients copy"
- Bind brand colours with hex codes — "the logo text 'ACME' in color #FF5733"
BFL's documented constraint on hex: codes work best when clearly associated with a specific object. A loose "use #FF5733 somewhere" produces inconsistent results. Bind every code to a named element.
When to use a Tool instead
A generation model is the wrong instrument for several common product jobs:
| Job | Right tool |
|---|---|
| Remove a stand, clamp or reflection | FLUX Erase (mask, no prompt) |
| Square shot needs to be a wide banner | FLUX Outpainting (target dimensions, no prompt) |
| Existing photo is slightly soft | FLUX Deblur (no prompt, no mask) |
| Garment on a model | FLUX VTO |
| Product in a new scene | FLUX.2 multi-reference editing |
The rule: if the product already exists correctly in a photo and you need to change the frame around it, that is a Tool. If you need the product somewhere it has never been, that is multi-reference editing.
Cost, honestly
Reference images are billed. Every one counts as at least 1 MP at your tier's rate, on top of the output. A four-reference edit on [pro] at 1MP output is the generation plus four reference charges — several times a plain text-to-image call.
For catalogue-scale work, that is still cheap against a photoshoot, but model it before committing: references are where the cost lives, not resolution.
FAQ
Can FLUX keep my exact product label? BFL documents preservation of branding, labels, shapes and materials when the product is supplied as a reference image.
How many product references can I use? Up to 8 at 1MP output on [pro], fewer at higher resolutions, 4 on klein.
Why does my label look wrong? Usually an unassigned reference or a low-quality source. Give each reference an explicit role and start from a clean, legible packshot.
How do I control brand colours? Hex codes bound to a named element, not floating in the prompt.
Should I generate at high resolution? Generally no — generate at 1MP with more references, then upscale.
How do I remove a prop from an existing shot? FLUX Erase with a mask, not a generation.
Is this cheaper than a photoshoot? Per image, dramatically. Just budget for reference charges rather than assuming per-image prices.
Bottom line
Product photography with FLUX is a reference discipline. Identity comes from images, context comes from the prompt, every reference gets a stated role, and text gets quoted and colour-bound.
Do the 9MP arithmetic before you design the workflow, and reach for Erase, Outpainting or Deblur when the job is fixing a frame rather than inventing one.
Start with the Flux 3 AI workspace to build the reference set your catalogue will run on.
Sources
- Product Consistency (BFL documentation) — what FLUX.2 preserves and documented example jobs
- Product Mockups (BFL docs) — describing product, surface, lighting and composition
- Multi-Reference Editing (BFL docs) — 9MP budget, reference counts, role assignment
- Official FLUX.2 prompting guide — optics, text rendering and hex colour binding rules
- FLUX Erase (BFL docs) — prop and reflection removal
- FLUX Outpainting (BFL docs) — reformatting product shots
- FLUX Deblur (BFL docs) — restoring soft product photos
- BFL API pricing — reference-image billing rules
- FLUX.2 max model page — top-tier product and retexturing positioning
- FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — multi-reference and photorealism claims
Scope note: capabilities and limits as of July 30, 2026. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.
