Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.
👉 Join on Whop
💬 Join the community for support, free tools and early news on Discord
Both of these models read your prompt through a Qwen3-family text encoder instead of the old CLIP/T5 setup. That makes people assume one prompt style works for both. It doesn't. Copy a prompt straight from one to the other and you'll get a worse result than either model is capable of. This guide breaks down exactly what changes and why.
Everything said here about Krea 2 applies equally to the base checkpoint and to full finetunes built on it, since a finetune retrains the diffusion weights, not the text encoder or the sampling recipe. The prompting rules come from those two, so they carry over unchanged. What a finetune changes is the aesthetic a given prompt produces, not how you have to write that prompt.
Recommended prompts for this comparison, one written for each model:
Z-Image Turbo prompt:
an elderly street food vendor in his sixties with grey hair and a deeply lined, weathered face, stirring a wok of noodles at a crowded night market stall, steam rising thickly around his hands, strings of bare bulbs and rows of red paper lanterns glowing overhead, a narrow market alley with blurred neon signage and indistinct crowds in the background, photorealistic documentary photograph, 35mm lens, shallow depth of field, visible film grain, slight motion blur on the stirring hand, natural aged skin texture, muted warm color grade, humid night air, candid unposed moment, no readable textKrea 2 prompt:
an elderly night market vendor with grey hair and weathered hands stirring noodles in a blackened wok, steam rising around his chopsticks, strings of warm bulbs and red lantern glow overhead, other stalls softly blurred behind him, photorealistic documentary photograph, available light, natural skin texture, shallow depth of fieldThe two models at a glance
Z-Image Turbo (ZIT) comes from Alibaba's Tongyi Lab. It's a 6B model built on the S3-DiT architecture, a single-stream design, with Qwen3-4B as the text encoder. It runs in 8 steps at a CFG close to 1.0 and fits in 16GB of VRAM.
Krea 2 comes from Krea.ai. It's a 12.9B diffusion transformer trained from scratch, with Qwen3-VL-4B as the text encoder. That's the vision-language version of Qwen3, not the plain text version ZIT uses, which is part of why Krea 2 can read style reference images the way it does. It was trained natively up to 2048x2048. The Turbo checkpoint runs in 8 steps at CFG 0.0 in Krea's own tooling. If you're running it through the standard ComfyUI KSampler node, use CFG 1.0 instead; that node's guidance math treats a literal 0.0 as pure unconditional output and breaks the image. A slower Raw checkpoint exists too, around 52 steps at CFG 3.5, meant for LoRA training and variety rather than daily use.
Same generation of text encoder family, different architecture, different training recipe. That's the whole reason the prompting rules diverge.
Prompt length and density
This is the single biggest difference between the two.
Z-Image Turbo wants dense, long descriptions. Short prompts and comma-separated tag lists genuinely confuse it. Aim for a full paragraph, roughly 100 to 180 words. Write it like you're describing the finished photo to someone who can't see it. Cover the subject, the scene, the light, and the camera feel in one continuous block.
Krea 2 is flexible on length. A 5 to 20 word prompt works fine for pure exploration, since Krea 2 is built to interpret and fill gaps rather than demand precision. A 30 to 80 word prompt gives you a controlled image. Complex scenes with a specified palette can run 80 to 140 words. Short prompts don't confuse Krea 2 the way they confuse ZIT.
Run this exact prompt, unchanged, on both models. This is where Z-Image Turbo usually goes generic or unstable while Krea 2 handles it fine.
A woman reading a book by a rain-streaked window, soft grey afternoon lightSentence structure and ordering
Both models front-load the first thing you write, so lead with your subject on either one.
Z-Image Turbo: subject and action, then scene, then composition and camera, then lighting and atmosphere, then style, then a short quality tail if needed.
Krea 2: subject, setting, composition and camera, lighting, mood, medium and style, texture detail. If the medium itself is the point (a risograph poster, a woodblock print), lead with the medium instead of the subject.
CFG and negative prompts
Neither model wants a classic negative prompt the way SDXL does.
Z-Image Turbo runs at a very low CFG, close to 1.0. At that setting it effectively ignores negative prompts. Phrase everything as a positive constraint instead. Instead of "no blur," write "razor-sharp focus, crisp detail." Push CFG up toward SDXL-style values of 6 to 9 and you'll just get oversaturated, blocky output. This is a distilled model trained around its own low-CFG recipe.
Krea 2 Turbo runs at CFG 0.0 in Krea's own tooling, effectively no classifier-free guidance at all. That makes negative prompts even less useful than on ZIT. Only add one when you're troubleshooting a specific recurring artifact, and keep it short. Everything else belongs in the positive prompt as a stated preference, not an exclusion.
One workflow-specific catch: the standard ComfyUI KSampler node doesn't accept CFG 0.0 the same way. Its guidance formula reads 0.0 as pure unconditional output, so the prompt gets ignored and the image breaks. On stock KSampler, use CFG 1.0 instead, which is that node's version of "no extra guidance." Krea-aware custom samplers, and Krea's own tools, handle 0.0 fine.
Weighting and emphasis
Drop the old (word:1.3) syntax on both models. Neither encoder reads it that way, and pushing weights too far breaks the image instead of emphasizing anything.
To emphasize something on either model, restate it in different words or pick a more specific word. "Oxblood" beats "red," "sodium vapor" beats "orange light."
Style control
This is where Krea 2 pulls ahead in built-in tooling. It ships with nine official style LoRAs, each with its own trigger phrase (monochrome ink wash style, rainy window style, vintage tarot style, and six more), plus mood presets that add both positive keywords and negative avoid-terms, style stacks that combine two or three presets, and style reference images with adjustable strength.
Z-Image Turbo has no equivalent official style layer. Its look comes almost entirely from the prompt itself, or from whatever community fine-tunes and LoRAs you add on top.
Recommended settings
Z-Image Turbo: 8 steps, CFG close to 1.0. Fits in 16GB VRAM.
Krea 2 Turbo: 8 steps, CFG 0.0, mu 1.15. On the standard ComfyUI KSampler node use CFG 1.0 instead, since that node breaks on a literal 0.0. Generate at full resolution since it was trained up to 2048x2048; generating small and upscaling wastes what the checkpoint can do. FP8 quantization runs around 12GB VRAM, BF16 around 24.8GB.
Short, uppercase text, since that's what both models handle best. Run the same prompt on both to see which one holds the text.
A small chalkboard sign outside a coffee shop that reads OPEN TODAY, propped against a brick wall, morning light, photorealistic street photographyQuick reference
Architecture Z-Image Turbo: S3-DiT, single-stream, 6B. Krea 2: diffusion transformer, 12.9B.
Text encoder Z-Image Turbo: Qwen3-4B, text only. Krea 2: Qwen3-VL-4B, vision-language.
Prompt length Z-Image Turbo: 100 to 180 words, dense. Krea 2: 5 to 140 words depending on the goal.
Steps and CFG Z-Image Turbo: 8 steps, CFG close to 1.0. Krea 2 Turbo: 8 steps, CFG 0.0 (1.0 on stock ComfyUI KSampler).
Negative prompt Z-Image Turbo: ignored at optimal CFG, use positive constraints instead. Krea 2: barely functional at CFG 0.0, troubleshooting only.
Weighting syntax Z-Image Turbo: not supported, restate instead. Krea 2: not supported, restate or use LoRA strength instead.
Style control Z-Image Turbo: prompt and community LoRAs only. Krea 2: nine official style LoRAs, mood presets, style stacks, style reference images.
Native resolution Z-Image Turbo: not officially specified past what fits in 16GB VRAM. Krea 2: trained up to 2048x2048.
The practical takeaway
If you're switching a prompt between these two models, don't copy and paste. Rewrite it. Expand it into a dense paragraph for Z-Image Turbo, or trust Krea 2 to work with something shorter. Same encoder family, same "write sentences, not tags" philosophy, but a very different appetite for detail and a very different toolkit for controlling style.