Regi Ellis
e915afb65f
- Add .github/workflows/publish_action.yml for Comfy registry - Update README: ComfyUI Manager install (recommended), Beta badge + warning - Minor docs tweaks
2.0 KiB
2.0 KiB
IllustriousVAEEncode
Converts your existing image into the AI's working format (latent). Like scanning a painting so the AI can understand and modify it.
What does it do?
- Takes any image and converts it to the AI's internal canvas format
- Enables you to start from existing artwork instead of noise
- Preserves the structure and composition of your original image
- Essential for img2img workflows, inpainting, and style transfer
Inputs
- pixels: Your source image (from Load Image, previous generation, etc.)
- vae: The encoder model (usually loaded with your main model)
- mask: (Optional) Select specific areas to regenerate (for inpainting)
- grow_mask_by: Expand masked area for smoother blending
Outputs
- LATENT: The encoded canvas ready for sampling/modification
Artist Tips
- Perfect for "variations" - encode your image then sample with low denoise (0.3-0.5)
- Higher denoise = more changes, lower = stays closer to original
- Use mask for selective editing (fix hands, change background, etc.)
- The VAE quality affects how well details are preserved
When to Use
- Style Transfer: Encode photo → Sample with art style prompt
- Variations: Make similar versions of existing art
- Inpainting: Fix or change parts of an image
- Upscaling: Encode → Upscale latent → Sample → Better details
Common Settings
- No mask: Full image gets processed
- With mask: Only masked areas change (great for corrections)
- Denoise 0.3: Subtle variations, keeps most original
- Denoise 0.7: Major changes, uses original as guide
- Denoise 1.0: Complete regeneration (why encode then?)
Example Workflows
Quick Variation:
- Load Image → VAEEncode →
- KSampler (denoise 0.4) →
- VAEDecode → Similar but unique version
Style Transfer:
- Load Photo → VAEEncode →
- Add style prompt → KSampler (denoise 0.6) →
- VAEDecode → Stylized version
For the reverse process (latent to image), see IllustriousVAEDecode