ComfyUI Feedback Sampler

Work in Progress

A custom ComfyUI sampler for creating Deforum-style zoom animations through iterative feedback loop.

Example use with SDXLTurbo model

Features

• Feedback Loop Animation - Iterative sampling with zoom, pan, and rotate
• ControlNet Sequence Support - Automatically detects batched hint images and applies per-frame ControlNet conditioning with automatic looping
• LAB Color Coherence - Deforum-inspired color matching prevents chromatic drift
• Anti-Blur Sharpening - Maintains detail at low denoise values
• Noise Injection - Prevents stagnation (Gaussian/Perlin)
• Contrast Boost - Keeps colors vibrant
• Batch Output - All frames as sequence for video creation

Inspired by Deforum Stable Diffusion

Compatibility

Works with:

• Stable Diffusion 1.5
• SDXL
• Flux models
• Basically any model thats compatible with native Ksampler
• Turbo/lighting models are recommended for faster iteration

Installation

1. Clone into custom_nodes folder:

cd ComfyUI/custom_nodes/
git clone https://github.com/pizurny/Comfyui-FeedbackSampler

2. Install dependencies (optional, for sharpening and Perlin noise):

pip install -r requirements.txt

Or for portable/embedded Python installation (run in ComfyUI_windows_portable folder):

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\Comfyui-FeedbackSampler\requirements.txt

3. Restart ComfyUI

Find node at: Add Node → sampling → custom → Feedback Sampler

Quick Start

1. Connect model, CLIP conditioning, latent, and VAE (required)
2. Set parameters:
   • zoom_value: 0.005 (zoom speed, + = in, - = out)
   • iterations: 30 (frame count)
   • feedback_denoise: 0.3 (lower = more coherence)
   • color_coherence: LAB (prevents color drift)
3. Connect all_latents → VAE Decode → Video Combine

Key Parameters

Parameter	Default	Description
zoom_value	0.01	Zoom speed (-0.5 to 0.5)
translate_x	0.0	Horizontal pan in pixels/frame (-100 to 100). Positive = camera pans right
translate_y	0.0	Vertical pan in pixels/frame (-100 to 100). Positive = camera pans down
angle	0.0	Rotation in degrees/frame (-180 to 180). Positive = counterclockwise
border_mode	zeros	Fill for out-of-bounds areas (zeros/border/reflection)
iterations	5	Number of frames
feedback_denoise	0.3	Strength per frame (0-1)
color_coherence	LAB	Color matching (None/LAB/RGB/HSV)
noise_amount	0.02	Prevents stagnation (0-1)
sharpen_amount	0.1	Anti-blur strength (0-1)

Recommended Settings

Smooth Animation:

feedback_denoise: 0.25-0.35
zoom_value: 0.001-0.005
color_coherence: LAB
sharpen_amount: 0.15-0.25

Fast Zoom:

feedback_denoise: 0.35-0.45
zoom_value: 0.01-0.02
color_coherence: LAB
sharpen_amount: 0.1-0.2

Pan Flythrough:

zoom_value: 0.005
translate_x: 5-15
angle: 0
border_mode: reflection
feedback_denoise: 0.3-0.4

Spiral Zoom:

zoom_value: 0.02
angle: 1-3
translate_x: 0
border_mode: border
feedback_denoise: 0.3

Slow Rotate:

zoom_value: 0
angle: 0.5-2
border_mode: reflection
feedback_denoise: 0.25-0.35

ControlNet Sequences

The Feedback Sampler automatically detects batched ControlNet hint images and applies them per-frame:

1. Load your ControlNet image sequence (e.g. using Load Image Batch)
2. Connect the batched images to your ControlNet Apply node as usual
3. Connect the ControlNet conditioning to the Feedback Sampler's positive input
4. The sampler will use one hint image per iteration, looping automatically if iterations > frames

This enables guided animations where each frame follows a different ControlNet reference — useful for motion-guided zoom animations or style sequences.

Troubleshooting

• Black frames: Connect VAE to node
• Blurry: Increase sharpen_amount to 0.2+
• Color bleeding: Use color_coherence: LAB
• Too much change: Lower feedback_denoise to 0.2-0.3

License

MIT License - See LICENSE

---

Example workflows included:
Feedback_Sampler_Flux.json
Feedback_Sampler_SD1.5.json
Feedback_Sampler_SD1.5_custom_startframe.json
Feedback_Sampler_SDXLturbo.json