← Blog
✎blog post

Krea Left the Pipeline, and One Post Still Says Otherwise

An open steel drive bay on a dark bench with one empty slot and a clean rail, cold teal light and a thin amber accent on the rail edge, empty space above the bay.

Krea 2 Turbo was slower than Z-Image Turbo in our 4 November picture table, and on a warm strip Z-Image looked better, so we removed Krea 2 and deleted its weights, which makes the 23 September background result impossible to rerun here.

Part 3 of 4 in Image models in a real pipeline. Previous: Z-Image at Four Steps Kept the Structure and Lost the Texture · Next: Fifty Stills and Four Clips for One Window Cleaner

We removed Krea 2 from our picture pipeline on 4 November and deleted its weights, 26 GB plus a 419 MB four-step file. Our post of 23 September says the pipeline uses Krea 2 for final backgrounds. That sentence is now wrong. The post's result, two wins in three for Krea on a background with empty space, can no longer be rerun on this machine.

The table that decided it

We timed five picture engines at 1024 pixels on one street prompt, cold start and warm. Cold is the first picture after loading, warm is the next one.

Engine Cold seconds Warm seconds
RealVisXL Lightning, 4 steps 4.0 2.0
Z-Image Turbo int8, 9 steps 20.3 8.0
RealVisXL V5, 34 steps 18.0 10.0
Krea 2 Turbo, fast preset 16.0 10.0
Krea 2 Turbo, 8 steps 38.9 16.0

We looked at the warm strip of pictures and chose Z-Image as the default. Krea 2 was removed, and its weights were deleted the same day.

The removal rests on that strip, which is a judgment, and on the timings, which are the only measurement in the record. We did not score the layout of each engine on a shared prompt set.

Two Krea numbers that do not agree

The September post reports Z-Image at 7.5 seconds and Krea at 16.5 to 17.5 seconds for a 1024 square picture, so Z-Image was about 2.2 to 2.3 times faster. It also reports a Krea background at 2048 by 1152 taking 40 to 44 seconds for eight steps.

The November table shows Z-Image at 8.0 seconds warm and Krea 2 Turbo at 10.0 seconds warm on its fast preset. The gap there is 25 percent, not two and a half times. The two tables used different settings and warm-up states, so neither corrects the other. But a reader who takes the post's ratio as the current speed difference is wrong.

What the 2-of-3 result was

The September post used one nursery prompt and three seeds at 1024 pixels. Krea won two of the three on layout, light and variety, and Z-Image won one on plant detail. The post says a review recorded the score but not which seed went which way. Its own falsifier asks for a blind reviewer. The body never says the 2-of-3 count was blind, so we cannot claim it was.

The post also records how each model failed. Z-Image ignored the requested empty third. Krea added black plastic trays under the terracotta pots. Those observations are the useful part, and they do not depend on the weights.

What we did not do

We did not test the empty-third prompt on Z-Image at four steps. The four-step post in this series measures speed and texture, not layout. So the one property that Krea won on is untested for the model we kept.

We have not edited the 23 September post. A correction line on it is a separate change, and it should say the pipeline sentence is out of date, not only that the speed comparison was made at other settings.

How to prove this wrong

The claim that Krea 2 won 2 of 3 cannot be rerun on this machine without the weights. A rerun needs the same model files, the saved prompt and seeds 1001, 2002 and 3003, and a reviewer who does not know which model made which picture. If Krea wins fewer than two of three in that test, the September result does not hold. The removal claim is simpler to check: the weight files are absent from the models folder, and the table's timings can be rerun for Z-Image alone.

The open question is layout. Does Z-Image keep the empty space when the prompt asks for it at the settings we now use? That is the question the removal leaves open.

Provenance

  • /home/mario/Brain/ComfyUI-v0.43.0/2026-11-04-picture-engines.md, lines 11 to 19 (the timing table, the Z-Image default, and the Krea 2 removal and weight deletion).
  • /home/mario/Projects/web-apps/amarbaroHQ/webstack/amarbaro.com/content/blog/krea-2-wins-the-background.md, lines 1 to 67 (the 23 September front matter, the 2-of-3 result, the 7.5 and 16.5 to 17.5 second timings, the 40 to 44 second batch, and the stale pipeline sentence at line 57).
  • /home/mario/Brain/ComfyUI-v0.43.0/2026-11-04-zimage-2s.md, lines 7 to 31 (the Z-Image four-step and schedule settings used as the current default).

Comments

No comments yet.

Log in to comment.