Showreel
15 seconds. The same character, in motion.
Twelve scenes, one person
Studio, street, office, night, movement. Every image comes with its own metadata file — prompt, seed, model, sampler, steps, resolution, GPU, render time and peak VRAM. Available on request for every single image.
Identity here is carried by reference-image conditioning from a single canonical portrait — no LoRA. This is stated so plainly because it changes what the method can and cannot do: see Limits.
Consistency in motion
Two approved clips. One motivated action per clip, a calm camera, a clear room — a rule that cost several discarded takes.
The 4-step accelerator LoRAs were tested and rejected: roughly six times faster, but they choke the motion so much that the motion prompt has no effect. These clips run at the full 20 steps with cfg 3.5. The A/B test is documented.
How it is made
Technology
- FLUX.2 Klein 4B (distilled) — base and distilled weights are both Apache-2.0, checked against the publisher's API, not assumed
- ComfyUI 0.35.0, native nodes only — no custom-node dependency that breaks elsewhere
- Wan 2.2 I2V A14B fp8 for image-to-video
- Identity via ReferenceLatent from one canonical portrait
- LoRA training: ai-toolkit, rank 32 / alpha 32, lr 1e-4, adamw8bit, EMA 0.99
What you receive
- The workflow as API JSON, saved from a real run — not a rebuilt demo
- One metadata file per image: prompt, seed, model, sampler, steps, resolution, GPU, time, VRAM
- With a trained LoRA: dataset and captions, trigger word documented
- Contact sheets for sign-off — and the rejected takes, if you want to see them
Reproducibility is the real deliverable. A beautiful image that cannot be recreated next month is worth little; a seed, a workflow and a documented method are worth a lot.
Measured performance
Measured on the 12 approved images and 2 clips — VRAM sampled every 0.5 s, render time per image by the clock. Not modelled and not rounded in my favour.
| Scene | Resolution | Time s | VRAM MiB | Avg W |
|---|
| Settled, resolution already loaded | 5.0 s per image |
|---|---|
| Median over a mixed 12-scene set | 9.6 s per image |
| First image at a new resolution | 12–27 s |
| Complete 12-scene set, pure GPU time | 136 s |
| RTX 3060 12 GB, same workflow | 21.7 s per image |
| 5 s clip, 832×576, full 20 steps | ~21 min |
| Usable clip including re-renders (~1 in 2) | ~40 min GPU time |
| Character LoRA, 2500 steps, 48 images, RTX 3090 | ~56 min |
The re-render rate is stated on purpose. Roughly every second clip needs a second attempt — so a usable five-second clip costs about forty minutes of GPU time, not twenty. Quoting the optimistic number would only push the shortfall into your schedule.
Limits
What this pipeline cannot do yet — before you ask. Anything not listed here as measured counts as unproven.
Open weaknesses, honestly
- Strong facial expressions drift. Reference-image conditioning holds identity well in calm shots. In a hearty laugh the face wanders — rounder cheeks, fewer freckles, a different smile. A character LoRA was trained for exactly this and did not solve it: tested at three strengths, identity still did not hold. The cause is known — the training set had almost no images with strong expression — and the fix is a new dataset, not different training settings. Until that is done and approved, calm shots are what is reliably delivered.
- No dedicated face-lock module. PuLID and InfiniteYou are not built in here. Whoever needs that in the workflow gets integration work — nothing already proven.
- No pose control. ControlNet is not set up. Image direction runs through prompt and reference, which is less precise than a pose rig.
- Video times apply only to the RTX 5070 Ti. The 3090 finished a LoRA training but has no measured video numbers; its measurements for image editing, 4K upscaling and restoration are under Services. For the 3060 there are image numbers, no video numbers.
- Second character in progress. The dataset is ready, the LoRA is training. Nothing of it appears here before it is finished and approved.
Capacity
Extrapolated from the measured single times, with a 25 % reserve for corrections and failed attempts. Explicitly extrapolated, not measured over a full week — that would be a different statement.
| Service | Per unit | Per GPU day |
|---|---|---|
| Stills, mixed scenes, one character | 9.6 s | The GPU is not the bottleneck |
| 5 s clips, usable after re-renders | ~40 min | 9 clips |
| New character LoRA incl. dataset | ~2–3 h | 2–3 characters |
For stills the GPU is not the bottleneck — a complete 12-scene set is 136 seconds of compute. The real work is direction and sign-off, and that is where the quality comes from.
More than characters. The same hardware also runs image editing, 4K upscaling, video restoration, 3D from a single image and synthetic training data. Each of these is documented with a measured number and its limit: see the services.
Working together
If this fits, write a short note — ideally with what is listed on the right.
Klarverk
IT support & online marketing
Nipoon Kawser Kalam · Zirndorf, Germany
kontakt@klarverk.de
Small business under §19 UStG (German VAT exemption) — invoices are issued without VAT. The work stays on hardware in Germany; nothing goes to an outside GPU service.
What I need to quote cleanly
- Your workflow JSON, if you have one
- References or a canonical portrait of the character
- Scene list and desired output sizes
- Whether the result must be reproducible later — that changes the method