TRUE GRID
Welcome
The big model comparison grids for Krea2 models.
This overview is not meant to collect best-in-class highest quality examples but same seed, same config generations to see which sampler / scheduler combinations work well.
⚠️ Under construction: all six models have their grids, examples and dynamic pose strips, and the LoRAs their examples. More notes on prompting follow. ⚠️
About Krea2
Krea2 is prompted in natural language only: describe the picture in full sentences instead of listing tags.
You generate on Turbo and train on Raw. Turbo is the fast version for making images, Raw is the base that LoRAs and fine-tunes are trained on. A LoRA trained on Raw is then used on Turbo.
Turbo type models run at 8 steps and CFG 1, which disables the negative prompt.
The official model is a general-purpose one, the fine-tunes and merges listed here are aimed at furry stuff.
Krea 2 is a 12B parameter model by Krea.ai, released as open weights in June 2026. It reads the prompt with a large text encoder, Qwen3-VL 4B. A prompt names the character and species, the pose, the clothing, the setting, the light and the style, all in sentences. Some of the fine-tunes below also understand booru or e621 tags, their tables say so.
The two versions in numbers:
Turbo: distilled from Raw, 8 steps at CFG 1. Most checkpoints are built on it.
Raw: undistilled, 28-50 steps with CFG enabled. Krea does not recommend it for normal generation.
On a checkpoint that was retrained a lot, a LoRA made for plain Krea 2 can behave differently or not work at all.
The checkpoints are big: about 24 GB in full precision (bf16), about 12 GB as fp8 or int8 and about 7 GB in 4 bit. The text encoder qwen3vl_4b and the VAE qwen_image_vae are separate files, both are in the Comfy-Org/Krea-2 repository on Hugging Face.
Forge Neo runs Krea 2 with its krea preset. In ComfyUI the text encoder is loaded with the CLIP type krea2.
The sampler grids below use five schedulers: Beta, Simple, Normal, SGM Uniform and Karras. The others in Forge Neo are left out:
- Uniform, DDIM, Turbo, KL Optimal, FlowMatchEulerDiscrete: untested. On Krea2 they work out to almost the same curve as Simple.
- Exponential, Polyexponential: untested. Like Karras they are made for the much wider noise range of SDXL and will most likely blur the image the same way.
- Align Your Steps, Flux2: untested. Both are tuned for other models.
- Linear Quadratic, Bong Tangent: tested, both break Res Multistep, DPM++ 2M and DPM++ 2M SDE.
Default settings for Krea 2 Turbo, as set by the krea preset of Forge Neo. The Turbo type models below share them unless their table says otherwise:
| Type | Data |
|---|---|
| sampler | Euler |
| schedule | Simple |
| gen | Steps: 8 |
| cfg | 1.0 |
| shift | 1.15 |
| size | 1024x1024 up to 2048x2048 |
The grids and the dynamic pose strips below were generated with these settings at 1024x1280, without a negative prompt, using the int8 convrot file of each model where one exists and the fp8 file otherwise.
Krea 2 Turbo (v1.0)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2732656?modelVersionId=3072332 |
| download int8 convrot | https://huggingface.co/Comfy-Org/Krea-2 |
| based on | Krea 2 Turbo (12B parameters, distilled from Krea 2 Raw) |
| precision | int8 convrot |
| official info | Steps: 8, CFG: off (0.0 in Krea's own code, 1.0 in Forge Neo and ComfyUI), Shift: 1.15, Size: 1k to 2k, up to 2048x2048 |
| notes | Turbo type model! The negative prompt has no effect at CFG 1. |
| Sampler / Scheduler grid | CFG / Steps grid |
|---|---|
CFG 1.0 / 8 steps  |
Euler / Simple  |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Dynamic Pose
Kroma (v0.3.1 turbo OPD)
| Type | Data |
|---|---|
| download huggingface | https://huggingface.co/lodestones/Kroma |
| download int8 convrot | https://huggingface.co/jakethessmake/Krea-Kroma-v0.3.1-Turbo-OPD-ComfyUI-Quants |
| based on | Full fine-tune of Krea 2, distilled to Turbo speed |
| precision | int8 convrot |
| official info | Steps: 8-12, CFG: 1.0-1.5, Shift: 1.15 |
| notes | Turbo type model! The author publishes it in full precision only (24.5 GB). The int8 convrot file is a conversion by another user and the one the grids here were made with. The model card does not say what it was trained on. |
| Sampler / Scheduler grid | CFG / Steps grid |
|---|---|
CFG 1.0 / 8 steps  |
Euler / Simple  |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Dynamic Pose
liquidmix_krea2_turbo (v2.11)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2739202?modelVersionId=3346486 |
| based on | Merge on Krea 2 Turbo |
| precision | fp8 |
| official info | Sampler: ER SDE / Euler a, Steps: 8, CFG: 1.5-1.8, Size: 1024-1536. The author recommends his own variant of the Qwen-Image VAE, which is part of the download. |
| notes | Turbo type model! Experimental merge that leans towards illustration, with stronger non-human features on furry and creature characters. Weaker at photorealism than plain Krea 2. |
| Sampler / Scheduler grid | CFG / Steps grid |
|---|---|
CFG 1.0 / 8 steps  |
Euler / Simple  |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Dynamic Pose
Ratatoskr (Krea2 v2.46)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2803045?modelVersionId=3171881 |
| based on | Fine-tune of Krea 2 Raw, changed by the author to run in fewer steps |
| precision | fp8 |
| official info | Sampler: Euler, Scheduler: Simple, Steps: 20-25. No CFG stated. |
| notes | Trained on a synthetic dataset of anthro, furry and fantasy creatures. Not all Krea 2 LoRAs work with it. The grids here use 8 steps; the author's 25 looked slightly overbaked. |
| Sampler / Scheduler grid | CFG / Steps grid |
|---|---|
CFG 1.0 / 8 steps  |
Euler / Simple  |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Dynamic Pose
willysKrea2FurryMix (v0.4)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2918022?modelVersionId=3301288 |
| based on | Merge of Krea 2 fine-tunes and LoRAs |
| precision | int8 convrot |
| official info | Steps: 8-12, CFG: 1-1.5. Tags and prose both work, detailed prose gives the best results. Tested by the author up to 4096x4096, artifacts can start at 2560x2560. |
| notes | Turbo type model! Artist tags are not understood. |
| Sampler / Scheduler grid | CFG / Steps grid |
|---|---|
CFG 1.0 / 8 steps  |
Euler / Simple  |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Dynamic Pose
Wulver (v0.5)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2881657?modelVersionId=3361227 |
| based on | Full fine-tune of Krea 2 Raw, official Turbo LoRA merged in |
| precision | int8 convrot |
| official info | Sampler: Euler, Scheduler: Simple, Steps: 8, CFG: 1.0, Shift: 1.15, Size: 1024. Prompt format: sfw, @artistname, description in full sentences. 1,113 artist styles through @artistname. Booru and e621 tags also work. |
| notes | Turbo type model! The author says a CFG above 1 burns the image; in the CFG grid here it held up to 2.0 at 8 steps. Non-Turbo files are published as well: 52 steps at CFG 3.5. |
| Sampler / Scheduler grid | CFG / Steps grid |
|---|---|
CFG 1.0 / 8 steps  |
Euler / Simple  |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Dynamic Pose
⬆️⬆️⬆️To the top⬆️⬆️⬆️
Furry Loras
Furry Enhancer (Krea2 V2.1)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/310964?modelVersionId=3120301 |
| architecture | Krea 2 |
| official info | Weight 0.8-1. 2 megapixel resolution recommended. Trained on a mix from drawings to photorealistic images. No trigger word stated. |
| notes | Tested on Krea 2 Turbo at weight 1.0. |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
krea2 furry (krea2_furry_0716)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2731677?modelVersionId=3140157 |
| architecture | Krea 2 |
| official info | Weight 0.2-1.5, this version may need a lower one. Triggers: fursuit, realistic fursuit, kemono fursuit, toony fursuit, anthro, furry, kemono, digital painting, anime style. Sampler: ER SDE for more detail, Euler or Euler a for stability. |
| notes | About a quarter of its training images are fursuits. Tested on Krea 2 Turbo at weight 0.8. |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Delken Furry Realism Enhancer (v1.0 Krea 2)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2458864?modelVersionId=3289747 |
| architecture | Krea 2 |
| official info | Trigger: @delken. Adds realism and fur detail to a checkpoint, not meant as a realism model on its own. No weight stated. The author notes that the fur can look matted and that it knows few species. |
| notes | Tested on Krea 2 Turbo at weight 0.8. |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
Eleptor's Furry Anthro Lora (Anima/Illustrious/Pony) (Krea)
| Type | Data |
|---|---|
| download civitai | https://civitai.com/models/2540770?modelVersionId=3242356 |
| architecture | Krea 2 |
| official info | Weight 0.2-0.4 for subtle enhancement or species correction. Trigger: (furry anthro SPECIES:1.4). Negative: white background. |
| notes | Tested on Krea 2 Turbo at weight 0.8, without the trigger and its prompt weight. |
Examples
| western cartoon | realistic | 3D animated film | oil painting |
|---|---|---|---|
 |
 |
 |
 |
⬆️⬆️⬆️To the top⬆️⬆️⬆️
Xeno443