Filip Ronaj · character pipelines Test your character$120

One face.Every frame.Scored.

I build ComfyUI pipelines that keep your AI character the same person across outfits, light and locations. Every delivery includes the test that proves it.

Test your character$120 Move the loupe over any frameTap any frame to inspect it

20 frames from 1 reference image. Average match 0.69.
Two faces count as the same person from 0.36.

Frame 01, neon street, rain: face match 0.75
010.75
Frame 02, cafe, morning: face match 0.49
020.49
Frame 03, mountain ridge, sunrise: face match 0.63
030.63
Frame 04, studio, green backdrop: face match 0.92
040.92
Frame 05, rooftop, laughing: face match 0.52
050.52
Frame 06, beach, dusk: face match 0.83
060.83
Frame 07, snowy street: face match 0.72
070.72
Frame 08, recording studio: face match 0.59
080.59
Frame 09, kitchen, daylight: face match 0.64
090.64
Frame 10, desert road, convertible: face match 0.66
100.66
Frame 11, art gallery: face match 0.76
110.76
Frame 12, night train: face match 0.84
120.84
Frame 13, rainforest: face match 0.74
130.74
Frame 14, black and white, blinds: face match 0.62
140.62
Frame 15, spaceship corridor: face match 0.79
150.79
Frame 16, flower market: face match 0.70
160.70
Frame 17, boxing gym: face match 0.66
170.66
Frame 18, concert crowd: face match 0.45
180.45
Frame 19, office whiteboard: face match 0.81
190.81
Frame 20, library, lamplight: face match 0.64
200.64
The reference image: a front portrait of the character on a plain background
Reference

Every frame, ranked.

The same twenty frames, placed by how closely each face matches the reference. Three other characters sit on the left for scale.

  • 0.69the average match across all twenty frames
  • 0.10the closest a different person gets
  • 0.45the lowest, frame 18 in a concert crowd, and still the same person

How a job runs.

Four steps. Every hand-over comes with the scores.

  1. 1

    You send the character.

    Three to five reference images or a written description, the style you want, your GPU or cloud setup, and the scenes you need first.

  2. 2

    The character test.

    Twelve fixed test prompts: new angles, outfits, light and expressions. Every image is scored against your reference, and each prompt gets a pass, drifting or rebuild verdict.

    2 days · $120

  3. 3

    The pipeline.

    A ComfyUI workflow tuned to your character. A LoRA is trained only when the reference-based workflow does not hold the face, and you are told when that happens.

    4 days · from $390

  4. 4

    Review, then hand-over.

    I see every image and the report before you do. Anything under the same-person line or visibly off-model is rendered again, however it scores. You get the files, the settings and a guide.

    Full usage rights for your project

Three ways in.

Fixed prices. Every package includes the drift report.

Character test

$120

A drift report on your character concept and 10 on-model images.

Delivery
2 days
Images
10
Workflow file
Not included
LoRA file
Not included
Revisions
1
Test your character$120

Character pipeline

$390

A ComfyUI workflow tuned to your character, a trained LoRA when it needs one, 40 images and the drift report.

Delivery
4 days
Images
40
Workflow file
Included
LoRA file
When one is trained
Revisions
2
Ask about a pipeline

Content engine

$890

The pipeline, plus batch generation from a spreadsheet, an image-to-video workflow, setup on your GPU or cloud and a recorded walkthrough.

Delivery
7 days
Images
40
Workflow file
Included
LoRA file
When one is trained
Revisions
3
Ask about an engine

Extras: another character +$250 · delivery in 48 hours +$150 · 100 more images +$80

Questions.

Can the character look like a real person?

Only with that person’s written consent. Without it, the character is original.

Will it run on my computer?

I check your GPU first. 12 GB of VRAM is a practical minimum for most setups, and I adapt the pipeline to what you have. No GPU? Say so and I run it for you.

Do I own the images and the workflow?

Yes. You get the files with full usage rights for your project.

What does the drift report show?

Twelve fixed test prompts, such as a new angle, outfit, light or expression. Each image gets a face-match score against your reference and a pass, drifting or rebuild verdict, so weak spots are fixed before final delivery.

Do you always train a LoRA?

No. A LoRA is trained only when the reference-based workflow does not hold your character. When one is trained, you get the file and I tell you why it was needed.

Who checks the images before I see them?

I do. I look at the full set and the report, and anything under the same-person line or visibly off-model is rendered again, however well it scores. The number supports the eye; it does not replace it.

Send me your character.

Three to five reference images or a short description, the style you want and your GPU or cloud setup. The drift report comes back in two days.

Prefer to hire through a platform? Find me on Upwork, Fiverr, Contra and Freelancer.com.