ComfyPilotComfyPilot
模型库 / LoRA / Krea 2 CS Flat Palette Anime
Krea 2 CS Flat Palette Anime

Krea 2 CS Flat Palette Anime

3061750

LoRA 18 次下载 更新于 2026-11-23

底模:Krea 2

styleflat coloraninekrea2

下载模型 在源站打开

文件清单 线路:镜像(hf-mirror.com) · 切换

下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。

权重 krea2-cs-flat-palette-anime-v1.safetensors 448 MB 下载

说明

Krea 2 CS Flat Palette Anime

作者:cloudseaside

About A style LoRA for Krea 2 that pushes renders toward flat-color anime illustration: clean cel shading, crisp edges, and a bright, poster-like palette. Detailed backgrounds stay intact (cafés, city streets, stages, overgrown ruins), but painterly or photographic rendering is replaced with flat color fields and graphic light. Trigger word (optional): flt_style Recommended strength

  • 0.75 – 1.0: the sweet spot. Flat colors and lifted, clean shadows, and the composition from your prompt is kept.
  • 0.5: a light touch. Keeps more of the base model's lighting and depth.
  • 1.25 – 1.5: very strong style, but outfit and hair colors begin to drift and the composition can shift.

Samples Each sheet shows the same seed and prompt at strengths from 0 (base) up to 1.5, so you can see what the LoRA does at each strength.

Tip: pair it with a 2.5D prompt The LoRA works especially well with semi-realistic prompts. Describe the image as "a semi-realistic anime illustration with a 2.5D look" and add realistic lighting and material cues (rim light, fabric sheen, depth of field, cinematic color grading). The LoRA then flattens the backgrounds and palette into a pop, poster-like look, while the character keeps some depth and gloss.

  • 0.25 – 0.5: the best balance for 2.5D. The fabric sheen and soft lighting stay, while the background turns into flat color blocks and halftone patterns.
  • 0.75 – 1.0: the same prompt becomes fully flat anime.
  • See the solo stage singer preview for the full range, from 2.5D at 0 to flat at 1.5.

Tested settings

  • Krea 2 Turbo (int8), 8 steps, CFG 1, sampler er_sde / scheduler simple
  • Works with both Danbooru-style tags and long natural-language prompts
  • Tested resolutions: 848×1536, 864×1520, 1008×1296, 992×1328, 1152×1152

Training

  • 42 images, 30 epochs, 1,260 steps, batch 1
  • Network dim 32 / alpha 32, LR 1e-4 (constant), AdamW8bit
  • bf16, fp8 base, 1024px bucket

示例图