ComfyPilotComfyPilot
模型库 / LoRA / CineVivid
CineVivid

CineVivid

3059975

LoRA 31 次下载 3 喜欢 更新于 2026-11-21

底模:Krea 2

filmlightingmoodystylecinematickrea 2

下载模型 在源站打开

触发词

cinevivid

文件清单 线路:原始(huggingface.co) · 切换

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

权重 cinevivid_v1.safetensors 218 MB 下载

说明

CineVivid

作者:sansiro1

AT A GLANCE

  • Strength: 0.75 for the full look, about 0.55 for calmer portraits and everyday scenes.
  • Trigger word: cinevivid (optional, the style also works without it).
  • Tested on: Krea 2 Turbo (fp8 scaled), Euler, Simple, 8 to 10 steps, CFG 1.
  • Loading: diffusion model only. In ComfyUI use LoraLoaderModelOnly; no text encoder LoRA is needed.

CineVivid is a style LoRA for Krea 2. It pushes an image toward a cinematic look: stronger directional light, deeper contrast, more depth and a richer feel for materials such as wet asphalt, satin, brass, glass, steam, wool and skin. It changes the look, not the subject, so it works on realistic film stills as well as on painted, engraved or illustrated scenes, on portraits, interiors, landscapes, sci-fi, fantasy and still life. WHAT YOU SEE IN THE EXAMPLES

  • Single images are made with the LoRA only. The comparison posts show the same prompt and the same seed without and with the LoRA, base on one side and CineVivid on the other.
  • The strength series post shows one prompt and one seed at base, 0.55, 0.75 and 1.0.
  • Every image carries its generation data, so you can open it, copy the prompt and try it yourself.

STRENGTH

  • 0.55: calmer and closer to the base look. Good for portraits, daylight and documentary-style scenes.
  • 0.75: the recommended default for the full cinematic look.
  • 1.0: mostly more contrast, harder shadows and detail pressure. Treat it as an effect, not as a quality boost.

PROMPTING

  • Plain descriptive English sentences work best. Example: "A full-body noir view of a woman with short red hair in an emerald satin gown standing in a rain-soaked alley under red neon, wet cobblestones, steam rising."
  • Every training caption starts with "cinevivid,", so you can put it at the start of the prompt. In my side-by-side tests (same prompt, same seed, the word at the start, at the end or left out) the images changed only in small details, not in style.

SETTINGS USED FOR THE EXAMPLE IMAGES

  • Checkpoint: Krea 2 Turbo (fp8 scaled)
  • Text encoder: qwen3vl 4b (fp8 scaled), type krea2
  • VAE: qwen image vae
  • Sampler: Euler, scheduler Simple, 8 to 10 steps, CFG 1
  • Sizes: 1024x1536, 1536x1024 and 1024x1024
  • LoRA strength: 0.55 to 0.75 (1.0 only in the strength series)

GOOD TO KNOW

  • It is a style LoRA. Hands, complex actions and exact object counts still depend on the base model.
  • At strength 1.0 faces look more weathered and complex actions (hands, kneading, pouring) become less plausible. At 0.55 to 0.75 portraits stay closer to the base look.
  • Portraits can come out older than the prompt asks for. If you want a younger face, state the age clearly and use the lower strength.
  • The look leans toward higher contrast. For very bright or airy images, lower the strength and describe the light you want in the prompt.
  • Tested on Krea 2 Turbo (fp8 scaled). Other Krea 2 checkpoints, precisions and samplers have not been verified.
  • The trigger word can occasionally show up as lettering on signs in the picture.

TRAINING Trained with Civitai's trainer on Krea 2 (Base): rank 32, 10,000 steps over 20 epochs, 236 images. This release is epoch 17. The dataset is 236 captioned images in portrait, landscape and square formats: fantasy, cyberpunk, steampunk and noir scenes, period interiors, cityscapes and landscapes, plus portraits and figure studies. Every caption starts with the trigger word and describes subject, setting, light and materials in plain English sentences. If you post images made with CineVivid, please link the resource so others can find it. Feedback and your own settings are welcome in the comments. Read the full write-up: https://civitai.com/articles/39009/cinevivid-v1-from-8-loras-to-one-almost

示例图