A decision framework for your specific situation

First, what layer of the stack are you building at? If the answer is 'we are building a model labModel labA company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google.,' the rest of this section does not apply to you and you have a different and harder business than the one described here. If the answer is 'we are building an application platformApplication platformA product that wraps underlying models in user-facing creative tools and workflow, capturing the value of usability and brand. Examples include Krea, Leonardo, and OpenArt. on top of existing models,' continue.

Second, who is your customer? Be specific. 'Creators' is not a customer. 'Independent fashion brands selling on Shopify who need product photography for their seasonal lookbooks' is a customer. The specificity matters because it determines everything downstream, what models you need, what workflow you need, what price you can charge, what moatMoatA durable competitive advantage that protects a business from being copied or undercut, such as workflow depth, proprietary data, or brand and community. you can build.

Third, what is the customer's alternative if your product does not exist? If the alternative is 'they would hire a $5,000 photographer,' you have a great business. If the alternative is 'they would use Midjourney or Krea or any of fifty competitors,' you have a hard business. The cost of the alternative determines what you can charge and how big the market is.

Once you know what you are building, the order of decisions matters. Decide them in this sequence:

1. Which base model(s) are core to your product? This decision is downstream of your customer's needs, for product photography, FLUX or Imagen. For illustrative work, FLUX or Midjourney. For Chinese-language text, Qwen-Image. For video, this gets more complicated and depends on cost tolerance. Pick one or two primary models and design around their specific strengths.

2. Build vs buy at the inference layer? At the early stage (no customers or fewer than 10,000 monthly active users), use an inference platformInference platformA company that turns raw rented GPUs into pay-per-generation APIs, hosting model weights and handling scaling so developers do not manage the hardware. Examples include fal, Replicate, and Modal., fal for media-specific work, Replicate for the broadest model selection, Modal for custom inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. logic. Defer the build decision until you have actual scale to justify it.

3. Workflow vs single-shot generation? If your customer needs consistent characters, branded styling, or any kind of multi-step refinement, you need a workflow tool. Either build on ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. (more powerful, more operational overhead) or on a hosted workflow platform (less powerful, less overhead). Single-shot generation is becoming a commodity feature; workflow is where the differentiation lives.

4. LoRA strategy? Per the LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. section earlier, decide whether to support no LoRAs, LoRA training as a service, BYO LoRA upload, or a marketplace. The right answer depends on whether your customers need character/style consistency or whether they are doing one-off generation.

5. Pricing model? Subscription with cap, pure credit, or hybrid. Subscription with cap is what most consumer products converge to. Credit pricing is what most developer-facing products use. Hybrid is what most prosumer products use. Pick the one that matches your customer's purchasing pattern.

6. Content moderation? Have a plan from day one, even if the plan is 'co-founder reviews flagged content for 30 minutes per day.' Do not skip this step. The first content moderation crisis is when most products discover they should have built moderation infrastructure six months earlier.

7. Legal and compliance? Understand your jurisdiction. Have terms of service that include indemnificationIndemnificationA vendor's promise to cover a customer's legal costs if the product triggers a lawsuit, offered by some licensed-data model providers as protection against copyright claims on generated output. you can defend. Know the IP exposure of the models you are using. If you are serving enterprise customers, consider commercial indemnification clauses. If you are serving the EU, plan for the AI Act.

None of these decisions are permanent, you can change them later, but each one has real cost to change once you have customers depending on the choice. Making them deliberately at the start is much cheaper than making them by default and then discovering you need to migrate.