The studio is where the curtain comes back. Reference sheets, the training setup, and the character-consistency method that keeps Charlie himself across every photo. Honest craft, written for the curious.
There are a dozen ways to do each of these, and the good ones change every few months. What does not change is the job, and what it costs you to hand that job to somebody else's servers. Run as much of it on your own machine as your patience and your hardware allow.
The one you open daily. What matters is not the model, it is whether it will take a character reference and give you the same face twice. Hosted tools are quicker to start. A local one costs you a setup afternoon and then costs nothing per image, forever.
A LoRA is a small file that teaches a model one specific face. There are many ways to make one and they all work. This is the piece to own outright. A trained character on a model you hold is the only version of her that cannot be repriced, rate limited or withdrawn.
Cloning is cheap now and the hosted ones are very good. Local text to speech has caught up further than people expect. The question to ask is who holds the voice profile, because a voice is harder to move than a look.
Every one of these tools is expensive, cloud only, and better than it was last month. Treat it as the part you rent rather than own, keep clips short, and build them from stills you already control so the character survives the trip.
Generated or licensed, both are fine. The only real requirement is that you can prove you may use it, which rules out most of what is easy to grab.
Reference images, palette, wardrobe, the rules she does not break, and a row for every published asset with its prompt, seed and reference set. A spreadsheet does this. This is the part that decides whether the character holds together, and it is the part everyone skips.
Why local wherever you can bear it. Not privacy, though that is real. It is that a character is an asset, and an asset living inside somebody else's API is one you are renting. Prices move, access changes, models get retired. The work you did to make her consistent goes with them.
None of this needs the best tools. It needs the same face, the same dog, the same light, week after week, and a written note of how you got there. That is the whole craft. The brand names are the easy part and they will all be different next year.
Here's a representative prompt from a recent post. Annotated below.
editorial photograph, woman in oat linen and a white cotton tee walking a golden retriever along a quiet headland path, australian coast, low golden-hour sun behind, eucalyptus on the right, salt haze, shot on 35mm film, kodak portra 400, shallow depth of field, --ar 4:5 --style raw --cref [maya-ref] --sref [coast-style]
Notice three things: cref is locking the character (Maya's face/hair), sref is locking the look (coastal palette + light), and kodak portra 400 is doing 70% of the colour science.
A free sampler from my prompt library, written in Midjourney syntax with Flux variants noted. The structure carries over to whatever you use. Reference images included.
The method in full: choosing identity anchors, culling the reference set, caption discipline, and the rules that hold temperament.
From prompt to final frame: the path I take for a typical morning post. Includes the asset-record template, so every image has its prompt, seed and references on file.
Iteration counts, rejected outputs, and the discipline of knowing when an image is right.
32 annotated prompts, the reference method that stops a character drifting, and where each image belongs.
Character consistency for animals. An animal is not a second character, it is a second reference competing with your first, and every decision follows from that.
Both together as The Studio Starter, AU$39 rather than AU$48. Nine PDFs, 109 pages, plus a short guide to the one rule that reverses between them.