Every face below is an original synthetic person — an ULTRA-trained LoRA (4,500+ step, rank-64) built by the INKFLOW studio for AI video production. Commission your own: you get the .safetensors model, a verified reference pack and a perpetual, royalty-free commercial license — pay once, use forever, owe nothing per use. $48. Nothing trains until you pick your face from 3 candidates.
Tried the rest? Now try the best — free.
We are that confident you will like it: the first ULTRA model is on us — complete file, full spec, commercial licence and all. If we are wrong it cost you nothing. We do not think we are wrong.
Not a trial, not watermarked, not cut down — the same ULTRA LoRA file we sell, trained 4,500+ steps at rank 64. Dariusz Kwiatkowski-Osei, an original synthetic person, yours to use commercially forever with no fee and no attribution. We would rather you judged the quality than took our word for it.
close-up
turning
over shoulder
laughing
low angle
full body
One model, six angles — including the ones that usually break a character LoRA. This is what you are downloading.
The free download is a genuine “try before you buy” sample, provided as-is. Paid LoRA / pack downloads are non-refundable once downloaded (a downloaded model can’t be returned) — which is exactly why the sample is free, so you can test first. Full Terms & Refund Policy.
No signup or card. Email is optional — add it only if you want the download link plus a one-time £49 pack offer. Check the rank yourself from the file you download — the snippet is below, and in the readme.
Aisha Okafor
ULTRA LoRA (4500-step, rank-64). Identical face in every shot, every episode.
World Pack: £29 at launch, list £49 — and it comes free with the Cast Pack. Cast Pack: £49 (about $62) for every actor on the shelf — list value $399 — with the whole World Pack included.
Buy it right now — no signup, no account
One click to Stripe, downloads on your screen within the hour (usually minutes). Perpetual royalty-free commercial licence on everything.
Everything listed here — shelf models now, in-training models the moment they verify at rank 64 — is part of the pack you buy today. Future themed collections may be sold separately.
Watch a complete film made with this catalogue
A full music-video film, every face and place from the pack: 14 ULTRA LoRAs — 7 characters, led by Aisha Okafor and Nico Rossi, and 7 trained sets & props. The same faces hold from first frame to last — that is the product, on screen. It is also a book trailer: the story is adapted from the ice-hockey romance “Checking the Rules”, written by this studio — the cast you can buy, acting the book we wrote. The film carries its AI-generated disclosure badge.
4 ULTRA LoRAs from this catalogue in one film — every one is in the pack below. AI-generated, disclosure badge burned in.
Watch Di Ride - Forge Line Brawl (hockey)
10 ULTRA LoRAs from this catalogue in one film — every one is in the pack below. AI-generated, disclosure badge burned in.
It isn’t just people — we train the world too
Vehicles, rooms and whole locations get the same ULTRA treatment, so the hero’s truck and the car park behind the rink look identical in every scene of every episode.Grab the one-image contact sheet of the whole world pack — save it, share it. Below is every trained set and prop in the pack today — everything listed - including items still in training - is included in the pack. Starter taste: smartphone, pickup truck and parking lot for £9, royalty-free.
We throw away most of the training set. That’s the point.
Every serious trainer will tell you the same thing: the training set decides the quality of a LoRA, not the settings. So that’s where we spend the effort. For each subject we generate around 40 candidates, embed them with DINOv2, measure each one against the set’s own centre, and discard whatever drifts. Twenty to thirty-six usually survive. If what’s left isn’t coherent enough, the trainer refuses to run at all — we’d rather spend the GPU hours again than sell you an average model.
Only then do we train deep — 4,500+ steps at rank 64, where community defaults sit nearer 1,500–2,500 at rank 16–32. Depth only helps once the data is clean, which is why the curation gate comes first. Together they hold a face through the hard cases: extreme angles, awkward lighting, motion blur, and a hundred shots without drift.
Don’t take our word for it — check the file yourself
We publish our training spec. Ask any other seller for their step count and rank before you buy. Ours is readable straight out of the .safetensors you receive, in one line. If it doesn’t say 64, you’ve caught us.
import json, struct
with open('actor.safetensors','rb') as f:
n = struct.unpack('<Q', f.read(8))[0]
hdr = json.loads(f.read(n))
print(min(v['shape'][0] for k, v in hdr.items()
if 'lora_A' in k or 'lora_down' in k)) # -> 64
Want something specific? Just ask a human.
Need a character or object trained to your exact brief? A bulk or studio licence? A book trailer, a music video, or a whole persistent world for your project? Want to see the spec, or have us prove a claim before you pay a penny? Email us and a person replies — no ticket system, no bot.
You receive plain .safetensors LoRA files. Load them in ComfyUI, Automatic1111 or any LoRA-capable pipeline, and add your own voice, motion and audio with whatever tools you already use. We sell the models, not a walled garden — and you can verify the rank-64 spec yourself, straight from the file header.
Questions buyers actually ask
Can I use these commercially?
Yes. Every purchase (and the free sample) carries a perpetual, royalty-free commercial licence: pay once, use forever, no per-use fee and no attribution required.
What do “4,500+ steps” and “rank 64” actually mean?
Training depth and model capacity. Community LoRAs typically train 1,500–2,500 steps at rank 16–32; ours train 4,500+ steps at rank 64 on a curated set, which is what holds one face steady through extreme angles, hard lighting and motion blur. Both numbers are readable straight from the .safetensors header — the one-line check is published above.
Are downloads refundable?
Paid LoRA and pack downloads are non-refundable once downloaded — a downloaded model cannot be returned. That is exactly why the full free sample exists: test everything before you spend a penny.
What exactly is included when I buy?
Everything listed on this page at the time you buy - shelf models immediately, in-training models when they verify. Future themed collections (style worlds, regional packs) may be released as separate packs.
Can I verify the quality before paying?
Yes, completely. Download the free ULTRA model — the same file class we sell, no signup — render with it, and read the rank out of the file header with the snippet above. If it doesn’t say 64, you’ve caught us.
Not a promise — a delivery record: one continuous canon of 2,138,487 words across 14 volumes (773 chapters) has already been written, with a single continuity ledger holding it together — the first series alone runs 1,254,473 words and crossed the 900,000-word mark with continuity intact.
Full-length novels, series and technical manuals up to 900,000 words — any genre, written natively in 22 languages, delivered autonomously by the same engine that has already shipped a 2.1-million-word continuous canon with machine-audited continuity. Describe the book, pay, download; a typical novel arrives in under 24 hours.
Also on the menu: complete books up to 900,000 words, single continuous stories past 2,000,000 words as multi-volume epics (we’ve already delivered a 2.1-million-word canon), whole multi-volume series with machine-audited continuity, manuals and technical specs — written natively in 22 languages. AI agents: connect to our MCP server at api.inkflowstudio.org/api/mcp and call inkflow_free_samples or inkflow_capabilities.
