4.2 The Method Ladder: Which Technique at Which Stage

There are five ways to hold a face in a diffusion pipeline, and the field's advice is confused because people recommend the one they use rather than the one that fits your stage.

This chapter orders them. The order is not about quality — it is about how much identity signal each method injects, and what it costs you to deploy. Climb as far as your project needs and no further.

The ladder

Read the ladder as a progression in the bandwidth of the identity signal, which is the concept from 4.1. A seed carries no identity information. A reference image carries an enormous amount. A trained adapter carries that information in a form the model has internalized. Each rung injects more, and each costs more to build.

Rungs 0 and 1 are not throwaway

Operators skip straight to adapters and then wonder why their first week was chaos. The first two rungs are what make everything after them debuggable.

Rung 0 — locked configuration. A single prompt, a single seed, a single set of sampling settings, generating the same image repeatedly. This is not a production method. It is your control condition: it proves your environment is behaving, and it gives you the reference for what "the same" looks like in your pipeline.

Rung 1 — the reference plate with img2img. One canonical image of the persona's face, used as the starting image at a low denoise strength. At low strength, the identity survives because you never fully left the pixels. The cost is that pose, expression, framing, and lighting are also inherited — which makes this a repair tool and a variation tool rather than a production method.

The reference plate is the artifact to create deliberately. Chapter 4.4 makes it part of the dataset; chapter 4.6 makes it the standard every output is measured against.

Rungs 2 to 5: the training-free adapters

These inject a face embedding extracted by a recognition model. They need no dataset, they work with a single reference, and they produce a first result in an afternoon — which is why they are the correct starting point and the subject of chapter 4.3.

Two path recommendations, because the base family from 3.1 determines your options:

On SDXL. The mature route is IPAdapter FaceID or InstantID. Both are widely installed, both run on modest hardware, and the SDXL base's permissive license avoids the licensing trap from 11.4.

On FLUX-class bases. The native route is PuLID: it works directly with the family, uses less VRAM than the alternatives, and interferes least with the base model's aesthetic. It is the correct first stop on a FLUX pipeline.

If you do not know yet: start with IPAdapter FaceID. It is the most documented, the most widely installed, and — most importantly — it can produce a usable first asset today. [AS OF 2026-10]

What none of them give you is what 11.4 called portability. An adapter is a stack, not a file: adapter weights, a face recognition model, a set of custom nodes, and specific version couplings between them. Change your tooling and the stack must be rebuilt. Chapters 4.3 and 12.4 both come back to that.

Rung 6: why the LoRA is still the only portable method

The LoRA is a single file, typically tens to a couple hundred megabytes, that teaches the base model what the face looks like. Three properties make it the destination rather than a stop on the way:

It is a file you own. Trained on a permissive base, it is an asset you can back up, move between machines, keep if a vendor disappears, and license. Chapter 11.4 explained why that matters commercially; chapter 12.4 explains why it matters operationally.

It carries the most identity signal per compute. Identity is baked into the weights rather than injected per generation, so it does not compete with your prompt for influence and it does not require a recognition stack to be running correctly.

It composes. A LoRA can be combined with the edit model, with style adapters, and with conditioning, which is rung 7.

Its costs are equally clear: you need a curated dataset, a training run of two to seven hours on a rented GPU for roughly $2 to $14, and a GPU capability from chapter 3.3. It also bakes in whatever your dataset contained — a documented trap that 4.4 is largely about.

Rung 7: the hybrid, which is where this ends

The LoRA is a poor scene model. An edit model is a poor identity model. Put them in the right order and each covers the other's weakness:

1. LoRA holds the face. Every generation carries the trained identity.
2. The edit model handles the scene. Instruction-based editing changes outfit, location, lighting, and pose while the face is already established in the pixels.
3. The reference plate resolves disputes. When the two disagree, the plate from 4.1 decides.

The failure mode this avoids is the most common one in the field: using a single method for everything and concluding that consistency is impossible. It is not impossible; it is a division of labor.

The gate: reference path or training path?

This is the decision this chapter exists to force. Five inputs:

One thing to notice about the dataset row, because it inverts the standard advice. For a LoRA of a real person, the dataset is the hard part — you need twenty-five good photographs of somebody. For a synthetic persona, you can generate the dataset. You have an unlimited supply of candidates, which means the hard part is not acquisition. It is selection: producing a set that is consistent enough with itself to be worth training on. That is the whole content of 4.4, and it is why the training path is more accessible for a persona than the tutorials written for real people make it sound.

The honest default: start on the reference path, ship something, and move to the training path once your volume or your patience runs out. Both thresholds arrive, usually within the first two months of real production.

What the ladder does not solve

Two problems get blamed on identity methods and are not identity problems at all.

Wardrobe and set continuity. Nothing in this ladder keeps a jacket consistent across a month. That is chapter 5.4's reference system, and it is a different mechanism.

Brand-level coherence. The ladder keeps one face recognizable. It does not make the feed look like one account. Color, framing, caption voice, and posting rhythm do that, and chapters 5.5 and 5.6 cover the parts that live in production.

Knowing which problem you have saves you from a training run that cannot fix it.

Where this goes wrong

Starting with a LoRA because it sounds serious. You will spend a weekend training on a face you have not validated, and discover its problems after the run rather than before it.

Skipping rung 0 and blaming the adapter. Without a control condition you cannot tell an adapter problem from an environment problem.

Never creating a reference plate. Every rung above depends on it, including the training path.

Treating adapters as permanent infrastructure. They are a stack with version couplings. Plan the exit before you need it.

Using an adapter's output as your training dataset. That contaminates the set with the adapter's own artifacts and shortcomings. Train from raw generations and from the plate.

Expecting one method to do both jobs. Identity and scene are different problems. Rung 7 is the answer, and it is a division of labor rather than a single technique.

Reading the ladder as a hierarchy of sophistication. It is a hierarchy of cost. Stop wherever your volume says to stop.

The gate questions

1. Which rung are you on today, and what put you there?
2. What is your monthly published-asset volume, and which side of the threshold is it on?
3. Do you have a reference plate — one file, chosen on purpose?
4. If your current tool disappeared, would your identity survive it?
5. Which of the five gate inputs above would have to change for you to climb a rung?

What you leave with

Seven rungs ordered by identity bandwidth and deployment cost, the two path recommendations per base family, the argument for why the LoRA is the destination rather than a stop, the division of labor that makes consistency achievable, and five inputs that answer the only question that matters at your current stage: reference path or training path?