Skip to content

extended-fill-types.md

Latest commit

 

History

History
164 lines (105 loc) · 9.95 KB

extended-fill-types.md

File metadata and controls

164 lines (105 loc) · 9.95 KB
Raw

Extended Fill Types

Feature: Additional noise fill patterns via DazzleNodes integration

Overview

Smart Resolution Calculator includes 5 built-in fill types for empty image generation: black, white, custom_color, noise, and random. These are always available regardless of what other nodes are installed.

When dazzle-comfy-plasma-fast (included in the DazzleNodes aggregate pack) is detected, 4 additional DazNoise fill patterns are automatically added to the fill_type dropdown. No configuration is needed -- detection happens at node registration time.

DazNoise Fill Types

DazNoise: Pink

Brightness-biased noise using cube root transformation. Produces lighter, more visible noise compared to standard uniform noise. The cube root mapping shifts the distribution toward higher values, resulting in a predominantly bright, airy texture.

DazNoise: Brown

Extreme brightness-biased noise using double cube root transformation. Very light, highlight-heavy noise where most pixel values cluster near white. Useful when you want a nearly washed-out noise pattern with subtle variation.

DazNoise: Plasma

Organic cloud-like patterns generated via diamond-square recursive subdivision. Produces smooth, flowing gradients that resemble plasma or clouds. Good for natural-looking backgrounds, organic textures, or as a starting point for artistic generation.

DazNoise: Greyscale

Monochrome noise where a single random value is generated per pixel and mapped identically across all three RGB channels. The result is pure grey-tone noise without any color cast.

DazNoise: Gaussian

Wide Gaussian noise centered on gray (mean=0.5, std=0.25). Produces a broader spread than the built-in noise fill type (which uses std=0.1). Generated via OmniNoise with Gaussian distribution mode. Good for backgrounds that need more variation than the tighter built-in Gaussian.

Custom Fill via fill_image Input

In addition to the dropdown fill types, a fill_image IMAGE input is available on the node. When an image is connected to this input, it overrides the fill_type selection entirely.

The connected image is scaled to match the calculated target dimensions. This means you can connect any noise generator (such as OmniNoise from DazzleNodes), any preprocessor output, or any image source as your custom fill.

Usage

  1. Connect any IMAGE output to the fill_image input on the node
  2. The fill_type dropdown is ignored when fill_image is connected
  3. The connected image is resized to the node's calculated width and height

This is useful for workflows where you want precise control over the fill pattern, or want to use a specialized noise generator that produces patterns not covered by the built-in options.

Seed Control for Noise Fills (v0.8.0+)

The SEED widget controls reproducibility of noise fills. When the seed toggle is ON, the RNG is seeded before generating the noise pattern, making the result deterministic.

Seed Widget State Behavior
Toggle ON, value >= 0 Fixed seed -- same noise pattern every time
Toggle ON, value = -1 Randomize each time (new random seed per queue)
Toggle ON, value = -2/-3 Increment/decrement last seed
Toggle OFF No RNG seeding; noise is non-reproducible; value passes through literally

Buttons (always functional in both ON/OFF modes):

  • Dice -- Set value to -1 (randomize each time when ON)
  • Lock -- Generate a new random seed value
  • Recycle -- Recall the last seed actually used (grayed when none). The last seed survives a workflow reload and an image drag-in (v0.12.3+): the recycle button is lit right after loading, and one click locks the widget to that seed.

Random-mode readout (v0.12.3+): while the widget is in randomize mode and a last seed is known, the value box shows that seed in a dimmed green-grey instead of Rnd: -1. The stored value is still -1 (the green tint and lit dice mean "re-rolls on the next queue"); the readout is the seed that made the current image, or the seed carried by the workflow or image you just loaded. Click to edit still shows the stored -1. When the workflow comes from an image (drag-drop or File > Open), the seed is read from the image's prompt metadata, the value Python actually received, so images made with any version, including 0.12.2 whose workflow block carries the previous run's seed, show and recycle their own seed (v0.12.4+). Loading a workflow JSON uses the saved property.

Note: When fill_type is black, white, or custom_color, the seed widget has no effect (these fills are deterministic regardless). When fill_image is connected, the seed is also irrelevant.

Latent Output Behavior (v0.8.1+)

The latent output depends on context:

Condition IMAGE output LATENT output
VAE + no image + noise fill Visual noise pattern (DazNoise, etc.) Raw torch.randn() Gaussian noise, seeded by fill_seed
VAE + image + transform mode Transformed image VAE-encoded transformed image
VAE + trivial fill (black/white/color) Solid color image Empty zeros latent
No VAE connected Fill pattern image Empty zeros latent

The latent dict includes a use_as_noise: True flag when noise fill is active. This flag is intended for downstream sampler integration (see below).

Caching: Noise image generation and latent results are cached. If the seed, fill_type, and dimensions are unchanged between runs, the cached result is reused (saves ~10s for expensive patterns like Plasma).

What fill_type Controls

  • IMAGE output: fill_type controls the visual pattern (Plasma blobs, Gaussian noise, etc.). Use the IMAGE output to preview or for img2img workflows at low denoise.
  • LATENT output: fill_type does NOT affect the latent when no image is attached. The latent is always seeded torch.randn() Gaussian noise — the same distribution diffusion models expect. The fill_seed widget controls which random noise pattern the latent contains.

In other words: changing fill_type from Plasma to Gaussian to Random produces different IMAGE outputs but identical LATENT outputs (given the same seed). The fill_type is a visual tool; the seed is the latent tool.

Why Not VAE-Encode the Noise Pattern?

VAE-encoded noise lives on the VAE's learned latent manifold — a structured space that diffusion models interpret as "partially denoised images." When used as starting noise for sampling, the model tries to denoise this structured input rather than generating from the prompt, producing abstract art instead of prompt-adherent images.

Raw torch.randn() noise is what diffusion models are trained to denoise. Using it produces proper prompt-adherent generation, just with a different seed than the sampler would normally use.

Spectral Blending (v0.8.2+)

When blend_strength > 0, the fill_type noise pattern's spatial structure is injected into the latent noise via FFT-based spectral blending. This allows the noise pattern to influence the composition of generated images while maintaining the Gaussian statistics diffusion models require. See the Spectral Blending Guide for algorithm details, empirical thresholds, and workflow patterns.

Image-to-Noise Fusion (v0.10+)

With the image_purpose widget set to img2noise, a connected input image replaces the fill_type pattern as the spectral blend source. The image's spatial structure shapes the noise composition. See the Image Purpose Guide for details.

Sampler Integration (Experimental)

The use_as_noise latent flag enables downstream samplers to use SmartResCalc's noise-filled latent as initial noise. This requires sampler-side support and does NOT work out-of-the-box with standard ComfyUI nodes.

ClownsharKSampler (RES4LYF)

A pre-patched copy of beta/samplers.py is included at docs/code/RES4LYF_beta_samplers.py. To use it, copy it over your existing file:

copy docs\code\RES4LYF_beta_samplers.py custom_nodes\RES4LYF\beta\samplers.py

The patches make three changes:

  1. Propagate the flag through latent_x (around line 336):

    if latent_image.get('use_as_noise', False):
    latent_x['use_as_noise'] = True

  2. Check the flag in the noise generation section (around line 549):

    use_latent_as_noise = latent_unbatch.get("use_as_noise", False)
if use_latent_as_noise and noise_seed == -2:
    noise = x.clone()

  3. Lower the seed minimum from -1 to -2 in INPUT_TYPES (both SharkSampler and ClownsharKSampler_Beta)

Set ClownsharKSampler's seed to -2 to activate noise passthrough. All other seed values use standard behavior.

Standard KSampler

Not supported. The standard KSampler ignores custom keys in the latent dict.

Detection and Availability

Condition Available fill types
DazzleNodes not installed black, white, custom_color, noise, random
DazzleNodes installed All 5 above + DazNoise: Pink, Brown, Plasma, Greyscale, Gaussian
fill_image connected Fill type dropdown ignored; connected image used instead

Detection is automatic. The node checks for the presence of dazzle-comfy-plasma-fast in ComfyUI's custom nodes at registration time. If found, the extended fill types appear in the dropdown. If not found, only the stock 5 fill types are shown -- no errors or warnings are generated.

Version History

  • v0.10.0: Image-to-noise spectral fusion via image_purpose widget
  • v0.8.4: Fixed seed serialization — workflow reload recovers actual seed used
  • v0.8.2: Spectral blending, noise caching
  • v0.8.0: Seed widget, VAE encoding of noise fills, noise caching, fill_type always visible, sampler integration docs
  • v0.7.0: Added DazNoise extended fill types and fill_image input