Services · measurements from 22 to 30 September 2026

Eight tasks, each measured on our own hardware.

Images, video, 3D and training data run locally here on three graphics cards. Every service below is a real test run with measured numbers and the limit we found along the way.

1,200images per hour, measured on all three cards
0.44 Whpower per image, graphics cards only
3 cardsRTX 3090 (24 GB), 5070 Ti (16 GB), 3060 (12 GB), used separately
0bytes of your data leaving the building
ApprovedInternal test = our own test run, no client work

One character, many scenes

Consistent characters · AI influencers · campaigns

The same person in every scene: face, freckles and hair colour stay stable, even across 36 images in a single run. Also from the front, three-quarter and profile, laughing, talking, listening, tired.

Measured
5.0 s per image (5070 Ti), 1,200 images per hour on three cards, 32 of 36 candidates approved in the production test; expression set: 28 training images, 8 of them with natural expression
Limit
Natural expressions hold in stills. A hearty laugh and expressions in video still change the face slightly; a first trained LoRA did not fix that, and the second version has not been judged yet
ApprovedFLUX.2 Klein 4B · Apache-2.0
Eight images of the same fictional woman: neutral, smiling, in profile, seated, full body, walking, in hard light and running
NOVA_01, fictional character · 8 of 32 delivered images: neutral, smile, three-quarter profile, seated, full body, walking, hard light, running
Twelve images of the same fictional woman with natural expressions and from different angles
Twelve images from the 28-image training set: laughing, smiling, talking, listening, tired, three-quarter and profile views · stills, not video

A still becomes a clip

AI video · social clips · product in motion

Five seconds of calm, motivated movement from a still image. The character stays recognisable, including in profile.

Measured
914 to 972 s per 5-second take at 576 × 832 (5070 Ti); earlier runs at 832 × 576 took about 21 minutes; roughly every second take is usable
Limit
Longer than 5 seconds, lip sync and large gestures not yet tested
Approved · internal testWan 2.2 · Apache-2.0
Office · 972 s render time
Rooftop · 922 s render time

Image editing, the person stays the same

Background · outfit · retouching · light

Six typical client requests on one image. Face, jewellery and posture stay unchanged; only what you asked for changes. Drag the slider to compare.

Measured
about 2 minutes per edit (RTX 3090), two variants to choose from
Limit
Tested on our own character; real client photos and logos not yet
Approved · internal testQwen-Image-Edit 2511 · Apache-2.0
Before After
BeforeAfter

Cut-outs in bulk

E-commerce · transparent PNGs · catalogues

Background gone, hair and handles stay clean. Built for hundreds or thousands of product images in a row.

Measured
about 1 second per image (5070 Ti), under 4 GB of graphics memory
Limit
Glass, transparent objects and very fine fur not yet tested
Approved · internal testBiRefNet · MIT
Cut-out hair on a checkered pattern
Hair edge, detail
Cut-out mug on a checkered pattern
Product with handle opening

Upscaling to 4K

Print · large displays · image archives

The model sharpens what is really in the image instead of inventing new faces. Left: plain enlargement; right: our result, same crop.

Measured
43 seconds per portrait image at 2624 × 3834, 79 seconds for a square at 4096 × 4096 (RTX 3090)
Limit
Portrait images above 4K showed faint seam lines, so we test your files first
Approved · internal testSeedVR2 · Apache-2.0
Plain enlargement, blurry
Plain enlargement
AI upscaling with eyelashes and freckles
SeedVR2, 4×

Video restoration

Old footage · family archives · smoother motion

We deliberately shrank a clip to a sixteenth of its pixels and compressed it heavily, then restored it and compared it with the original. On top of that, the frame rate was doubled for calmer motion.

Measured
67 s for 5 seconds of video (RTX 3090), no added flicker, SSIM 0.79 to 0.81
Limit
Fine details such as freckles are rebuilt, not recovered
Approved · internal testSeedVR2 · RIFE · Apache-2.0 / MIT
left damaged, right restored at 32 instead of 16 frames per second
One frame in four versions: damaged, enlarged, restored, original
damaged · plain enlargement · restored · original

3D model from one image

Web shop · AR · game assets

A product image becomes a rotatable 3D model. The shape is right from every side, including the unseen back. The four input images are themselves AI-generated; real product photos are next. Rotate the model with mouse or finger.

Measured
26 to 63 seconds per object (5070 Ti), delivered as GLB for web and AR
Limit
Surfaces still show white speckles; the texture-baking step is still missing
Approved · internal testTRELLIS.2 · MIT · Built with DINOv3

The rotatable 3D view needs JavaScript. Fixed views below.

Traffic cone, from a single image
Four objects: input image and four views of the 3D model
left the input image, right four views of the model

Synthetic training data

Object detection · segmentation · YOLO / COCO

Labelled images for rare or privacy-sensitive subjects. The labels are exact because we know where each object was placed. The green boxes are the delivered labels.

Measured
500 images, 1,338 objects with box and pixel mask; the 80 source images took about 10 minutes on two cards, the total composition time was not measured separately
Limit
In dark scenes the objects look pasted in; a trained detector has not been tested yet
Approved · internal testFLUX.2 Klein · own code
Twelve scenes with traffic cones and green detection boxes
Sample with the labels drawn in

Licences and limits

Every model was checked at its source for commercial use before we used it. We do not use models with a non-commercial or EU-excluding licence for client work.

FLUX.2 Klein 4BApache-2.0 · images
Wan 2.2Apache-2.0 · video
Qwen-Image-Edit 2511Apache-2.0 · editing
SeedVR2 7BApache-2.0 · 4K and restoration
BiRefNetMIT · cut-outs
RIFE 4.9MIT · frame rate
TRELLIS.2MIT · 3D
DINOv3DINOv3 licence · Built with DINOv3

All people and companies shown are invented and machine-generated. Measurements come from our own test runs between 22 and 30 September 2026; they apply to the settings described, not to every task.