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
IllustriousTriplePassSampler
A three-stage sampler for artists who want maximum control: composition, structure, and detail passes, each with their own settings.
What does it do?
- Runs three denoising passes: composition (global layout), structure (shapes), and detail (edges/textures).
- Lets you tune steps, CFG, denoise, samplers, and schedulers for each pass.
- Supports adaptive step/CFG/denoise tweaks for sharpness and stability.
- Version-aware: auto-detects EPS/VPred models and applies best presets.
Inputs
- model: Your UNet diffusion model (SDXL/Illustrious).
- seed: Random seed for reproducibility.
- positive/negative: Conditioning prompts.
- latent_image: Starting latent tensor.
- comp/struct/detail steps/cfg/denoise: Control each pass.
- sampler/scheduler: Choose your favorite sampler for each pass.
- progressive_cfg_decay: Gradually reduce CFG for stability.
- adaptive_steps/refinement: Auto-tune steps and sharpness.
- preserve_composition: Keeps the initial layout stable.
- model_version: Explicitly set model version or leave on auto.
- Version Preset Override: Force EPS/VPred presets if needed.
Outputs
- composition_latent: After composition pass.
- structure_latent: After structure pass.
- final_latent: After detail pass (ready for decode).
Artist Tips
- Use high comp denoise for wild layouts, then rein in with structure/detail.
- Try different samplers for each pass to experiment with style.
- Adaptive steps/refinement can help with sharpness in complex scenes.
- Combine with Smart Scene Generator for rich, multi-layered prompts.
- Use with Empty Latent for blank-canvas workflows or with outpaint for expansion.
Example Workflow
- Generate a latent with Empty Latent node.
- Use Smart Scene Generator for a creative prompt.
- Run TriplePassSampler with adaptive settings for cinematic results.
- Decode and color correct as desired.
For simpler workflows, see IllustriousMultiPassSampler.