Krea 2 Wins the Background
Krea 2 won the background task 2 to 1 because it followed the empty-space instruction. Z-Image stayed sharper and faster, so the choice depends on the job.
Z-Image Turbo at 768 pixels and four steps made a warm picture in 1.7 seconds on one RX 7900 XTX, down from 7.1 seconds at 1024 pixels and nine steps, and kept the structure of three test prompts.
Part 2 of 4 in Image models in a real pipeline. Previous: Krea 2 Wins the Background · Next: Krea Left the Pipeline, and One Post Still Says Otherwise
At 768 pixels and four steps, Z-Image Turbo made a picture in 1.7 seconds once the model was warm. At 1024 pixels and nine steps it took 7.1 seconds. The shop, park and face prompts kept their structure at the smaller setting. The fine texture did not survive it.
Before the sweep, a stall in our own client cost up to two seconds on every picture. The client asked the server whether a render had finished only once every two seconds. Cutting that poll to 50 milliseconds took the nine-step 1024 picture from 8.0 seconds to 7.1, with no change to the model.
We checked whether a faster kernel could help before changing settings. At 1024 pixels the model needs about 55 trillion floating-point operations per step. We measured 0.65 seconds per step, which is 85 trillion operations per second, or 69 percent of the card's 16-bit peak. Nine steps at the integer peak would still take about 2 seconds plus overhead. A custom engine could not reach the 1 to 2 seconds we wanted.
So the levers were steps, pixel count and client overhead. Those are the three we changed.
Warm seconds per picture, client side, after the poll fix:
| Steps | 1024 pixels | 768 pixels |
|---|---|---|
| 9 | 7.1 | 3.5 |
| 6 | 5.5 | 2.4 |
| 4 | 4.4 | 1.7 |
The server ran at 1.54 iterations per second at 1024 pixels and 2.83 at 768. Overhead outside the sampler was about 0.3 seconds at 768. Dropping from 1024 to 768 roughly halves the work per step, which is why the 768 column is close to half the 1024 column.
We looked at contact sheets of the shop, park and face prompts at four settings. At four steps and 768 pixels, the structure, composition and light held in all three. The loss was fine texture: produce clutter, leaves and skin pores went soft. Six steps added little for 0.4 seconds more.
These judgments were made by us, looking at the sheets. They were not blind, and they were not scored.
We then tried the noise schedule at four steps and 768 pixels, on the face and shop prompts with one seed. All arms took 1.7 seconds, except five steps at 2.1. A shift of 1.5 brought back skin wrinkles, stubble, jacket grain, plaster cracks and produce density on both prompts. A shift of 6 flattened the picture. The beta schedule re-composed it and washed the sky. Shift 1.5 is now the default.
Shift 1.5 is the one place where we chose a quality setting by eye and then made it the default. It rests on two prompts and one seed.
The default for zimage in our picture tool is four steps at 768 pixels with shift 1.5. Jobs that need the finer texture ask for nine steps at 1024. Three default jobs run without a size argument took 8.1 seconds cold and 1.7 seconds on each warm run.
We did not train a four-step distillation LoRA for Z-Image Turbo. The receipt found none on the Hub that fits, and training one would take about half a day of setup. That is not justified until a use needs the texture back.
A speed number at 768 compared with a speed number at 1024 is not a quality result. A quality claim from one prompt is not a default. The schedule choice rests on two prompts and a single seed.
Rerun tools/picbench.sh with the zsteps job file. Warm seconds should fall near 7.1, 5.5 and 4.4 at 1024, and near 3.5, 2.4 and 1.7 at 768. If they do not, the table is wrong. For the quality default, give a reviewer pairs of pictures from the two settings, with the settings hidden, and ask which keeps the structure. If the nine-step set wins most pairs, the four-step default is wrong for structure, whatever the speed.
The open question is texture. Three subjects are not enough to know which product photos lose too much at four steps.
/home/mario/Brain/ComfyUI-v0.43.0/2026-11-04-picture-engines.md, lines 9 to 19 (the picture-engine table and the default decision)./home/mario/Brain/ComfyUI-v0.43.0/2026-11-04-zimage-2s.md, lines 7 to 31 (the roofline, the poll fix, the steps and size sweep, the look, the schedule sweep, the default and the LoRA decision).
Comentarii
Niciun comentariu încă.
Conectare pentru a comenta.