On the same nursery prompt and three seeds, Krea 2 produced the more useful background composition in 2 of 3 comparisons, while Z-Image kept sharper plant detail.
Krea 2 won the background task 2 to 1. Z-Image was sharper, brighter, and faster, but it ignored the instruction that mattered most for a website hero: leave the left third soft and open. The result is a choice between composition and throughput, not a universal model ranking.
The empty third decided the comparison
The test used one nursery prompt and three seeds, rendered through ComfyUI at 1024 square for the direct side by side. The prompt asked for young Artemisia plants in a polytunnel, warm morning light, and an out-of-focus left third with room for text.
Krea won two of the three seeds on layout adherence, light, and variety across seeds. Z-Image won the third on clean plant detail. The review recorded the 2 to 1 score and its reasons, not which seed went which way, so we do not name them here; the six saved images let anyone redo the call.
The saved comparison is six images, not a claim derived from model names. The image sheet and individual outputs are in the ComfyUI A/B receipt set. At the same 1024 square size, warm, Z-Image took 7.5 seconds per image and Krea 16.5 to 17.5 seconds, so Z-Image was about 2.2 to 2.3 times faster. A later Krea background batch at 2048 by 1152 took 40 to 44 seconds per eight-step image.
The models fail in different directions
Z-Image's failure was compositional. It made the scene legible, crisp, and nearly the same across seeds, but it treated the negative space as optional. That is a bad default for a hero image with copy over it.
Krea's failure was object detail. It introduced black plastic nursery trays beneath the pots and made the Artemisia identity less certain. Its lighting and depth looked more photographic, but the scene needed a prompt correction to remove the trays.
That distinction matters. A model can be better at atmosphere and still be worse for a product claim. The next Krea prompt adds that the terracotta pots stand directly on the wooden bench.
Pick the model after naming the constraint
Choose the model for the failure you can fix in the layout, not the model with the nicest single frame.
What counts as a Krea win:
the requested empty region survives across seeds;
the hero can carry text without a crop rescue;
the lighting and depth support the page's visual hierarchy.
What does not count:
one attractive seed picked after the fact;
a quality claim made from different prompts;
a speed comparison where resolution and warmup differ.
The cost was detail and time
Krea cost sharper plant detail and a longer final render. Its standard eight-step 2048 by 1152 run took 40 to 44 seconds in the saved background batch. Z-Image remained the fast batch tool and the better choice for crisp detail shots.
That trade was acceptable for a hero image and not for exploring dozens of prompt variants, so for a while the pipeline used Krea for final backgrounds and Z-Image for quick fills and detail passes.
What we use now
Since 4 November 2026 the pipeline runs on Z-Image Turbo alone. Krea 2 left the stack: one model now covers backgrounds, detail shots, cover plates and stock pictures, and the time it saves goes into more prompts per job instead of slower renders. The quality arm is 9 steps at 1024 square in about 7 seconds; the default is 4 steps at 768 square in under 2. How we got Z-Image that fast is in A Z-Image picture in 1.7 seconds, and why a finished render is not yet a correct one is in Eight ways a ComfyUI job succeeds and is wrong. The composition lesson from this test still holds: Z-Image follows one clear composition per prompt, so the empty space for copy has to be described as part of the scene, not asked for as an instruction.
How to prove this wrong
Use the saved nursery prompt, seeds 1001, 2002, and 3003, and the same ComfyUI model files. Hide model names from the reviewer. Count only whether the left third remains useful negative space, then count plant fidelity separately. If Krea wins fewer than two layout comparisons, or if the plant-fidelity result reverses under the same rubric, this post's claim needs correction.
Provenance
One RX 7900 XTX, ComfyUI, Krea 2 Turbo fp8, Qwen3-VL text encoder, Qwen image VAE, eight steps, Euler, simple scheduler, CFG 1.0. Receipts: the saved A/B image set, the shared comparison workflow, and the background render log. The direct comparison was an instrument receipt at 1024 square, while the background batch was 2048 by 1152.
Four steps at 768 pixels cut a Z-Image picture from 7.1 seconds to 1.7. The sampler is already near the card's compute limit, so the gain came from steps and size, and the texture loss is the price.
Krea 2 was slower in our picture table, and Z-Image looked better on the warm strip, so we removed it. The post that praised Krea for backgrounds is now out of date, and its two-to-one result cannot be rerun on this machine.
The stills batch ran clean at 10 to 14 seconds per picture, then the upscale pass stopped on a missing file and 11 pictures never reached the output folder. The clips are timed, but nobody has judged them.
Komentarze
Brak komentarzy.
Zaloguj się aby skomentować.