ComfyUI Advanced Generation Pack (AGP)

ComfyUI Advanced Generation Pack (AGP)

Production-ready custom nodes for advanced diffusion workflows. Modular implementation grounded in proven techniques from leading ComfyUI repositories.

Installation

Copy the repository into your ComfyUI custom_nodes directory:

```bash cd /path/to/ComfyUI/custom_nodes git clone https://github.com/your-repo/ComfyUI-AdvGenPack.git ```

Nodes will appear in ComfyUI UI under "AGP/" categories after restart.

Module Overview

WAN Multi-Expert Sampling

Automatic expert switching optimized for Wan2.2 models and Lightning LoRAs.

Nodes:

• WAN Multi-Expert Sampler - Switch between high-noise and low-noise experts at optimal diffusion boundaries
• WAN Lightning Hybrid (3-Stage) - Triple-stage sampling with per-stage LoRA strength optimization
• WAN Boundary Optimizer - Detect optimal expert transition points using gradient/entropy analysis
• WAN Noise Expert Mixer - Blend expert outputs with configurable interpolation modes

Key Features:

• Adaptive boundary detection
• Per-stage LoRA scaling (early: 0.8x, mid: 1.2x, late: 0.6x)
• Multiple blending modes (linear, sigmoid, cosine)

Example Config: ```toml [wan.multi_expert] boundary_type = "adaptive" cfg_scale = 7.5 expert_weights = "1.0,1.0" noise_blend_mode = "sigmoid" ```

Frame Interpolation

RIFE/DAIN-style frame synthesis with optical flow and temporal consistency.

Nodes:

• Optical Flow Frame Interpolator - Generate intermediate frames using Lucas-Kanade or Horn-Schunck flow
• Temporal Consistency Enforcer - Maintain coherence across frame sequences
• Keyframe Extractor - Auto-extract critical frames from sequences
• Motion Prediction Sampler - Predict intermediate frames with easing functions

Example Config: ```toml [frame_interp.optical_flow] flow_method = "lucas_kanade" interpolation_mode = "flow_guided" blend_strength = 1.0 ```

Post-Processing FX

Real-time image effects for final rendering polish.

Nodes:

• Bloom & Glow FX - Threshold-based bloom with falloff (additive, screen, soft-light modes)
• Denoise Suite - NLM, bilateral, median, morphological denoising
• Sharpening Filter - Unsharp mask, high-pass, detail enhancement
• Chromatic Aberration - RGB channel offsets (radial, directional, swirl)
• Film Grain & Noise - Procedural grain (uniform, Gaussian, Perlin-style)

Example Config: ```toml [fx.bloom] bloom_threshold = 0.8 bloom_intensity = 1.0 bloom_size = 15 glow_strength = 0.5 bloom_mode = "screen" ```

Color Correction

Professional color grading and correction tools.

Nodes:

• Histogram Equalizer - Adaptive/global histogram stretching
• LUT 3D Loader - Apply 3D LUT color grades
• Curves & Levels - Point-based curve adjustment
• HSV Shifter - Independent hue/saturation/value control
• Color Grading Preset - Pre-built cinematic grades (cinematic, vintage, cool, warm, noir)
• White Balance Corrector - Temperature/tint correction with auto-detection

Example Config: ```toml [color.grading] preset = "cinematic" intensity = 1.0 temperature = 6500 tint = 0.0 ```

Model & LoRA Merging

Advanced checkpoint and LoRA merging techniques.

Nodes:

• Checkpoint Merger (DARE) - DARE-TIES merging with configurable sparsity
• Checkpoint Merger (Block-wise MBW) - Per-block weight blending for SD1.5/SDXL
• Checkpoint Merger (Advanced Multi) - Multi-model blending with layer schedules
• LoRA Multi-Merger - Batch merge N LoRAs with custom alphas
• LoRA DARE Merger - DARE-based LoRA combination
• LoRA Strength Scheduler - Dynamic LoRA weights across diffusion steps
• Checkpoint Mask Generator - Generate layer-wise weighting masks

Example Config: ```toml [merge.dare] sparse_factor = 0.95 rescale_mode = "vector"

[merge.blockwise] architecture = "sdxl" input_blocks_weight = 0.6 middle_block_weight = 0.7 output_blocks_weight = 0.5

[merge.lora_scheduler] schedule_type = "sigmoid" start_strength = 1.0 end_strength = 0.5 ```

Workflow Examples

Example 1: Wan2.2 Multi-Expert Sampling

``` ckpt_load("wan2.2-base.safetensors") → wan_multi_expert( steps=20, boundary_type="adaptive", cfg_scale=7.5 ) → wan_lightning_hybrid( early_lora_scale=0.8, mid_lora_scale=1.2, late_lora_scale=0.6 ) → vae_decode() ```

Example 2: Frame Interpolation + Post-FX

``` load_frame_sequence("input/") → optical_flow_interpolator( interpolation_factor=0.5, flow_method="lucas_kanade" ) → temporal_consistency_enforcer( consistency_strength=0.7 ) → bloom_glow_fx( bloom_intensity=1.2, bloom_mode="screen" ) → film_grain_noise( grain_amount=0.05, grain_type="gaussian" ) → save_images() ```

Example 3: Model Merging + Grading

``` ckpt_load("sd15-base.safetensors") ckpt_load("style-a.safetensors") ckpt_load("style-b.safetensors") → checkpoint_merger_blockwise( architecture="sd15", input_blocks_weight=0.6, middle_block_weight=0.7, output_blocks_weight=0.6 ) → lora_load("quality-lora.safetensors") → lora_strength_scheduler( steps=25, schedule_type="cosine", start_strength=1.0, end_strength=0.4 ) → ksampler(steps=25) → color_grading_preset(preset="cinematic", intensity=1.0) → vae_decode() ```

Technical Details

Architecture

• Modular Design: 7 separate node modules (WAN, Frame Interp, FX, Color, Merging, Utils)
• Utilities: Common functions, color space conversion, merge algorithms in utils/
• Type Hints: Full Python 3.8+ type annotations for IDE support

Performance Notes

• Optical flow operations CPU-optimized for speed (Lucas-Kanade, Horn-Schunck simplified implementations)
• Bloom/denoise operations support batched tensor operations
• Merging operations designed for efficient state dict manipulation

Compatibility

• ComfyUI: Latest versions (tested on 0.1.x+)
• Python: 3.8+
• PyTorch: 2.0+
• Models: SD1.5, SDXL, Wan2.2

Node Reference

Input/Output Types

• MODEL - Diffusion model checkpoint
• LORA - LoRA checkpoint dictionary
• IMAGE - Tensor [batch, height, width, channels] in [0, 1] range
• LATENT - Latent space tensor
• CONDITIONING - Prompt conditioning from CLIP encoders
• SIGMAS - Noise schedule for diffusion steps
• INT, FLOAT, STRING, BOOLEAN - Standard types

Common Parameters

• steps - Number of diffusion steps (1-500)
• cfg_scale - Classifier-free guidance scale (0-30)
• strength - Effect intensity (0-1 or 0-2 depending on context)
• schedule_type - Interpolation mode (linear, sigmoid, cosine, step)
• blend_mode - Blending operation (additive, screen, soft_light, overlay, etc.)

Troubleshooting

Nodes not appearing in UI:

• Ensure repo is in custom_nodes directory
• Restart ComfyUI completely
• Check console for import errors

Out of memory errors:

• Reduce bloom_size in Bloom FX
• Lower frame interpolation search area
• Split large frame sequences into batches

Inconsistent results with frame interpolation:

• Increase consistency_strength in Temporal Enforcer
• Use flow_guided interpolation mode
• Extract keyframes first with Keyframe Extractor

Citation & References

Grounded in techniques from:

• WanMoeKSampler (Wan2.2 expert routing)
• TripleKSampler (3-stage Lightning optimization)
• DareMerge (DARE-TIES merging)
• LoRA-Merger-ComfyUI (LoRA combination)
• RES4LYF (frame and image processing)
• ComfyUI-Extra-Samplers (additional sampling methods)

License

MIT License - See LICENSE file for details