I'm going to be upfront: we are an AI production agency, so take this guide with that in mind. But that's also why it's worth reading. We see the competitive landscape from the inside. We know which agencies are doing serious work and which ones are running generic prompts through Midjourney, charging brand rates, and calling it AI production. The gap between those two things is enormous -- in quality, in brand consistency, and in what you actually get for your money. These are the seven questions I'd ask if I were evaluating any studio, including ours.
Question 1: Do you do model fine-tuning for brand consistency?
This is the single most important technical question. Fine-tuning -- whether through LoRA (Low-Rank Adaptation), Dreambooth, or similar methods -- means the AI model has been specifically trained on your product, your brand colors, your packaging. Without it, every image generation is essentially a guess. The model has no memory of what your product actually looks like.
When we onboard a new client, we spend the first few days building a brand-specific LoRA. We shoot reference images, tag them carefully, train the adapter, test it against a validation set. That process takes 2-3 days and costs real money. Agencies that skip this step will tell you their results are 'consistent' but they're lying to themselves as much as to you. Ask specifically: do you train a custom model for my brand? How many reference images do you use? Can I see the training data we'd provide?
Question 2: Can I see the model outputs, not just final polished work?
Any good agency will show you their best work. What reveals actual capability is seeing the raw outputs before retouching and compositing. If an agency is hesitant to show you these -- if the portfolio only contains finished images with no production notes -- that's worth questioning. We show clients raw generation grids on request. The gap between our raw outputs and a competent post-production pass is much smaller than what you'd see from a generic prompt shop.
Also look at portfolio diversity. Do they show work across different product categories? AI models behave very differently for, say, glossy cosmetics packaging versus rough-textured food products versus fashion. An agency that only shows one type of work may only know how to do that one thing.
Question 3: What does your production workflow actually look like?
A proper AI production workflow has more human involvement than most clients expect -- and that's a sign of quality, not inefficiency. Our workflow for a typical product shoot campaign: strategic brief and art direction (human), reference photography or 3D asset preparation (human or vendor), LoRA training (human-supervised), generation runs with iteration (human-directed), selection and curation (human), compositing and retouching (human), final delivery. AI is doing the heavy lifting in generation but humans are making decisions at every other stage.
If an agency describes their workflow as 'we write the prompt and deliver the images,' be cautious. That might be fine for certain use cases -- social content, quick iterations, concept exploration. It's not fine for hero campaign imagery.
Question 4: What are realistic prices for different project scales?
| Project Type | What's Included |
|---|---|
| Quick concept exploration | 20-30 generated images, no fine-tuning, basic selection |
| Social content batch (monthly) | Brand-trained model, 40-60 final images, copy variants |
| E-commerce product catalog | LoRA training, 100+ product variants, background compositing |
| Campaign hero images (4-6 finals) | Full fine-tuning, art direction, retouching, rights clearance |
| Video content (10-15 seconds) | Kling/Runway generation, editing, sound, delivery |
Anyone quoting you for a 'fully custom AI campaign' is doing generic prompts. Anyone quoting for something in the middle tiers above is probably padding. These ranges come from our actual invoices. Markets vary -- Istanbul is cheaper than London -- but the ratios hold.
Question 5: Who owns the trained model weights?
This question will tell you a lot about how sophisticated an agency is. When we train a LoRA on your brand assets, the resulting model weights are technically a derivative of both the base model and your reference images. Ownership questions get complicated fast. Our standard contract specifies that clients own the right to use generated outputs commercially, we retain the trained adapter weights for active clients, and weights are transferable on contract termination. That's a reasonable structure. Some agencies don't think about this at all and have no contract language covering it.
Also ask about the base models they're using. Are they working with commercial-licensed models (Stable Diffusion with appropriate licenses, proprietary APIs like Midjourney or Adobe Firefly)? Or are they fine-tuning on models trained on scraped data without clear commercial rights? The legal landscape here is still evolving, but you don't want to build a campaign on assets with unclear IP status.
Question 6: Can you show me a failed project?
Not a question every agency will welcome. But the answer tells you a lot. We've had product categories that genuinely don't work well with current AI tools -- certain reflective jewelry, highly textured handmade ceramics, products where the physical material itself is the story. When we encounter those, we say so and recommend traditional photography. An agency that claims AI works for everything hasn't done enough volume to know its limits yet.
Question 7: What's your revision and approval process?
AI generation is fast, which creates a dangerous expectation: that revision cycles are also fast and free. They're not. Each revision requires re-running the model with adjusted prompts, re-selecting from a new output grid, re-compositing. Build time expectations into your brief. We typically allow two structured revision rounds in our base pricing and charge hourly for additional rounds. That's not us being difficult -- it's us being honest about where the real labor sits in an AI production.
Frequently Asked Questions
Is it worth paying more for a fine-tuned model versus generic AI outputs?
For anything going on your website, packaging, or paid ads -- yes, absolutely. Generic outputs will get your general vibe across but they'll never consistently represent your actual product. For early-stage concept exploration or internal ideation, generic is fine.
How long does it take to train a brand-specific model?
A LoRA training run itself takes a few hours of compute time. But the process of shooting or sourcing reference images, tagging the training set, running the training, validating outputs, and iterating typically takes 2-4 working days. Budget for that in your project timeline.
Can a small brand afford proper AI production?
Yes, but it requires some upfront investment that pays out over time. Training a LoRA once and using it for 6 months of content makes the per-image cost very low. The mistake is treating each project as a new engagement -- build a relationship with one studio and amortize the setup costs.
What red flags should immediately disqualify an agency?
No examples of actual client work (only stock-like demonstrations), inability to explain their technical process in plain terms, no contract language about IP ownership, and prices that are either suspiciously low or unjustifiably high. Also: agencies that promise AI can do everything. It can't.
We're happy to answer these questions about our own work. If you want to see our LoRA training process, our raw output grids, or our standard contract language before deciding anything, just ask. We'd rather have that conversation than win a client who has the wrong expectations.