Spaces:
Running
on
Zero
Usage
Enter a prompt and click Generate.
Prompting
Positive and negative prompts are embedded by Compel for weighting. You can use a float or +/-. For example:
man, portrait, blue+ eyes, close-upman, portrait, (blue)1.1 eyes, close-upman, portrait, (blue eyes)-, close-upman, portrait, (blue eyes)0.9, close-up
Note that ++ is 1.1^2 (and so on). See syntax features to learn more and read Civitai's guide on prompting for best practices.
You can also press the 🎲 button to generate a random prompt.
Arrays
Arrays allow you to generate different images from a single prompt. For example, [[cat,corgi]] will expand into 2 separate prompts. Make sure Images is set accordingly (e.g., 2). Only works for the positive prompt. Inspired by Fooocus.
Embeddings
Select multiple negative textual inversion embeddings. Fast Negative and Bad Dream can be used standalone or together; Unrealistic Dream should be combined with one of the others:
<fast_negative>: all-purpose (default)<bad_dream>: DreamShaper-style<unrealistic_dream>: realistic add-on
Styles
Styles are prompt templates from twri's sdxl_prompt_styler Comfy node. Start with a subject like "cat", pick a style, and iterate from there.
Scale
Rescale up to 4x using Real-ESRGAN (Wang et al. 2021).
Models
Each model checkpoint has a different aesthetic:
- lykon/dreamshaper-8: general purpose (default)
- fluently/fluently-v4: general purpose merge
- linaqruf/anything-v3-1: anime
- prompthero/openjourney-v4: Midjourney-like
- runwayml/stable-diffusion-v1-5: base
- sg161222/realistic_vision_v5.1: photorealistic
Schedulers
The default is DEIS 2M with Karras enabled. The other multistep scheduler, DPM++ 2M, is also good. For realism, DDIM is recommended. Euler a is worth trying for a different look.
Image-to-Image
The 🖼️ Image tab enables the image-to-image and IP-Adapter pipelines. Either use the image input or select a generation from the gallery. To disable, simply clear the image input (the x overlay button).
Denoising strength is essentially how much the generation will differ from the input image. A value of 0 will be identical to the original, while 1 will be a completely new image. You may want to also increase the number of inference steps. Only applies to the image-to-image input.
IP-Adapter
In an image-to-image pipeline, the input image is used as the initial latent. With IP-Adapter (Ye et al. 2023), the input image is processed by a separate image encoder and the encoded features are used as conditioning along with the text prompt.
For capturing faces, enable IP-Adapter Face to use the full-face model. You should use an input image that is mostly a face along with the Realistic Vision model. The input image should also be the same aspect ratio as the output to avoid distortion.
Advanced
DeepCache
DeepCache (Ma et al. 2023) caches lower UNet layers and reuses them every Interval steps:
1: no caching2: more quality (default)3: balanced4: more speed
FreeU
FreeU (Si et al. 2023) re-weights the contributions sourced from the UNet’s skip connections and backbone feature maps to potentially improve image quality.
Clip Skip
When enabled, the last CLIP layer is skipped. This can sometimes improve image quality with anime models.
Tiny VAE
Enable madebyollin/taesd for near-instant latent decoding with a minor loss in detail. Useful for development.
Prompt Truncation
When enabled, prompts will be truncated to CLIP's limit of 77 tokens. By default this is disabled, so Compel will chunk prompts into segments rather than cutting them off.