Krea 2#
Krea 2 is a 12-billion-parameter MMDiT model and the bot-wide default -- if you run
/imagine without touching model:, this is what renders your image. Three versions are
available: Krea, Krea (Quality) and Krea (Exp).
It takes prose, not tags. It accepts a reference image. It has no negative prompt at all.
In the picker they appear as one Krea 2 family option that expands to Krea -> Krea (Exp) -> Krea (Quality).
At a glance#
| Version | Steps | CFG | Sampler | Scheduler | Rec. steps | Rec. CFG | Base latent | Default batch |
|---|---|---|---|---|---|---|---|---|
| Krea (default) | 12 | 1 | euler |
simple |
8-16 | 1-2 | 1024x1024 | 2 |
| Krea (Quality) | 12 | 1 | euler |
simple |
8-16 | 1-2 | 1408x1408 | 1 |
| Krea (Exp) | 16 | 1.5 | euler |
simple |
12-24 | 1.2-2.5 | 1024x1024 | 2 |
All three share the family's default upscaler, 4x_NMKD-Siax_200k.pth, and the
Krea 2 Community License, which is what shows in the 🔍
Details view. None of the three ship quality tags, safe-content
guard strings, or a negative prompt.
The dropdown text says 8 steps
Both Krea and Krea (Quality) describe themselves as an "8 steps at CFG 1" recipe. The sampler actually runs 12 steps. The table above is what happens.
The three versions#
Krea#
The distilled turbo build, tuned to land its result in 12 steps. The balanced option, the bot-wide default, and the one to pick unless you have a specific reason not to.
| Architecture | MMDiT 12B (Krea 2), pre-distilled turbo INT8 ConvRot |
| Subtype | krea2, turbo_distilled |
| Prompting | Prose. Long descriptive sentences, read by a Qwen3-VL encoder. |
| Negatives | None. The graph zeroes the negative branch out entirely. |
| Steps | 12 (8-16) |
| CFG | 1 (1-2) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | NMKD-Siax_200k (4x) |
Good at: long prose prompts with spatial relationships and material description, rendered
text, reference-image conditioning through its vision encoder, and giving you two images per
/imagine to choose between. It accepts the Detailer, so faces,
eyes and hands can be repaired after the fact.
Weak at: exclusion and variety. There is no negative prompt at all, so the only way to keep something out of a frame is to describe a frame it does not fit in. Being pre-distilled, it also gives markedly less composition variety across seeds than Krea (Exp) does -- re-rolling the same prompt tends to return the same picture with the details shuffled. CFG stays at 1, so raising it is not a lever for prompt adherence. Danbooru tag prompts are the wrong shape entirely; use SDXL for those. And 12B at 1024x1024 batch 2 is a real load on one home GPU, so this is not the model to spam during a busy queue.
Krea (Quality)#
Same weights and the same 12-step recipe as Krea, sampled at a much larger latent and decoded without the compensating downscale. This is the "make it big" option.
| Architecture | MMDiT 12B (Krea 2), pre-distilled turbo INT8 ConvRot |
| Subtype | krea2, turbo_distilled |
| Prompting | Prose. Long descriptive sentences, read by a Qwen3-VL encoder. |
| Negatives | None. The graph zeroes the negative branch out entirely. |
| Steps | 12 (8-16) |
| CFG | 1 (1-2) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1408x1408 latent, which decodes to roughly 2816x2816 |
| Default batch | 1 |
| Default upscaler | NMKD-Siax_200k (4x) |
Good at: detail that survives at full size. Skin texture, fabric weave, distant architecture and small rendered text hold up here in a way they do not on the default version, because the 2x decode reaches the final image instead of being scaled back down.
Weak at: cost, at every stage. It is by a wide margin the most expensive thing you can ask this bot for: roughly 1.9x the latent area of the default version before the decode doubles the output. VRAM, render time, compression time and Discord upload time all go up together, and the batch is pinned to 1 for exactly that reason -- you get one image, not two. Everything Krea is weak at (no negatives, no CFG lever, no seed variety, tags are wrong) applies here unchanged, because it is the same weights.
Krea (Quality) returns roughly double the size you asked for
Krea 2 decodes through a 2x VAE. On the default version a compensating downscale brings the image back to the size you requested. On Quality that downscale is deliberately removed so the extra detail survives, which means a 1:1 render comes back at about 2816x2816, not 1408x1408. Budget the VRAM and the upload time accordingly. Batch defaults to 1 for the same reason.
Quality latents by aspect ratio (each roughly doubles at decode):
| 1:1 | 9:7 | 7:9 | 4:3 | 3:4 | 3:2 | 2:3 | 16:9 | 9:16 | 21:9 | 9:21 |
|---|---|---|---|---|---|---|---|---|---|---|
| 1408x1408 | 1600x1216 | 1216x1600 | 1600x1216 | 1216x1600 | 1728x1152 | 1152x1728 | 1920x1088 | 1088x1920 | 2112x896 | 896x2112 |
Krea (Exp)#
The raw, undistilled checkpoint with the turbo LoRA stacked back on at strength 0.6. Pick it when you are hunting for a composition and the default version keeps handing you the same one.
| Architecture | MMDiT 12B (Krea 2), raw INT8 ConvRot plus turbo LoRA at 0.6 |
| Subtype | krea2, raw |
| Prompting | Prose. Long descriptive sentences, read by a Qwen3-VL encoder. |
| Negatives | None. The graph zeroes the negative branch out entirely. |
| Steps | 16 (12-24) |
| CFG | 1.5 (1.2-2.5) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | NMKD-Siax_200k (4x) |
Good at: noticeably more composition variety from seed to seed than the pre-distilled build, which is the reason to pick it. It is also the only Krea version with a CFG range worth touching -- 1.2 to 2.5 is a real span, not the token 1 to 2 the distilled builds carry.
Weak at: resolution and speed. It only holds together at the 1024 latent, so the "render it bigger" move that Quality exists for is closed off here. At 16 steps it is a third more sampling work than the default version for the same picture. The turbo LoRA it stacks back on also occupies one of the family's 4 LoRA slots. No negative prompt, same as the other two.
Do not push Krea (Exp) past 1024
Krea (Exp) is only reliable at the 1024 latent. It warps at 2048, so do not push it with a large custom resolution.
Resolutions#
Krea and Krea (Exp) default to 1024x1024 and share this table:
| 1:1 | 9:7 | 7:9 | 4:3 | 3:4 | 3:2 | 2:3 | 16:9 | 9:16 | 21:9 | 9:21 |
|---|---|---|---|---|---|---|---|---|---|---|
| 1024x1024 | 1152x896 | 896x1152 | 1184x896 | 896x1184 | 1280x864 | 864x1280 | 1280x720 | 720x1280 | 1536x656 | 656x1536 |
Krea (Quality) uses the larger table in its section above.
How to prompt Krea 2#
Write sentences. Krea 2 reads your prompt through a Qwen3-VL text encoder, so it handles long descriptive clauses, spatial relationships and material descriptions the way a captioning model would. Comma-separated Danbooru tags are the wrong shape for it -- use SDXL for those.
| Thing | On Krea 2 |
|---|---|
| Negative prompt | Does not exist. The graph zeroes it out. Anything you put in negative: is discarded, so describe what you do want instead. |
| Quality tags | None are added. Your text is encoded verbatim. |
| Character names | Written in natural form -- Hatsune Miku, not hatsune_miku \(vocaloid\). See Character matching. |
| Prompt enhancement | Opt-in. Pass enhance: True on /imagine if you want the LLM to expand your idea. Unlike Ideogram 4, Krea does not turn it on for you. |
| CFG | Stays at 1 (1.5 on Exp). Raising it is not how you get more prompt adherence on a distilled model. |
Attach a reference image
The input_image: attachment on /imagine is wired straight into Krea 2's vision path.
Krea looks at the picture and conditions on it. This is not img2img -- there is no
denoise strength, and the source image is not the starting latent. Use it for style,
subject or palette reference.
LoRAs#
Five of the 91 LoRAs are Krea 2 compatible. Up to 4 stack at once, and on this family they patch the model only, leaving the text encoder untouched. See LoRAs.
That pool of five is small enough to be a real limitation. If restyling matters more to you than prompt handling, Z-Image has 44 compatible LoRAs and no slot limit at all.
Note
Turning your nsfw setting on while using Krea 2 quietly loads a content-filter-bypass
LoRA in addition to whatever you selected. It appears in the 🔍
Details credits chain like any other LoRA.
Hi-res fix on Krea 2#
Krea 2 ships its own hi-res fix recipe that is deliberately gentler than the global one:
| Setting | Global default | Krea 2 default |
|---|---|---|
| Denoise | 0.35 | 0.25 |
| Second-pass steps | parent steps x 0.75 | 6 |
The lower denoise exists because Krea 2 renders a coherent image in 12 steps; a 0.35 re-denoise at that fidelity rewrites detail you already wanted. If your hi-res pass is changing faces or losing the original composition, you have raised the denoise too far. The accepted range is 0.1 to 0.7.
The upscale factor slider runs 1.25 to 2.0 as normal.
Krea 2 in editing#
Two of the 15 /edit models are Krea 2 builds, both open to every
user (no VIP required):
| Edit model | Steps | CFG | What it is for |
|---|---|---|---|
| Krea 2 Edit Fast | 10 | 1.0 | Quick instruction edits on the turbo build |
| Krea 2 Edit Quality | 20 | 3.0 | The removal recipe (below) |
Use Quality to remove things
Krea 2 Edit Quality is specifically the "remove or delete salient content" recipe. Removing an object relies on negative-prompt image grounding, and that grounding only takes effect above CFG 1. Krea 2 Edit Fast runs at CFG 1, so it tends to simply re-render the subject you asked it to delete. If your removal keeps coming back, switch to Quality.
Both edit models accept steps 1-60 and CFG 0.5-7.0, force batch size 1, and take one input image only. Two advanced knobs are available:
| Knob | Range | Note |
|---|---|---|
| Reference boost | 0-20 | 4.0 is the recommended starting value |
| Grounding pixels | 384-1536 | Trained range is 384-768. Higher values help faces but leave the trained window. |
What does not work on Krea 2#
| Feature | Status |
|---|---|
| Negative prompt | Not supported at all |
| Quality tags | None shipped |
| Safe-content guard strings | None shipped, so your nsfw setting changes no prompt text here |
| img2img | Not the mechanism -- an attachment is a vision reference instead |
| Detailer | Supported |
| USDU upscale | Supported, but your prompt is replaced with a generic detail prompt for the tile pass |
| Hi-res fix | Supported, with the gentler defaults above |
Related#
- Model overview -- the full catalog and how the picker is organised
- Z-Image -- the other prose-prompted family, with unlimited LoRAs
- /imagine -- every parameter you can pass
- /edit -- the two Krea 2 edit models