Custom nodes are add-ons that plug into ComfyUI’s graph system to add missing functions (advanced samplers, IP-Adapter, ControlNet preprocessors, ESRGAN upscalers, AnimateDiff, utility math, batch tools). They let you:

• Build cleaner anime line art and cel shading pipelines
• Improve character consistency via LoRA/IP‑Adapter
• Add motion with AnimateDiff nodes
• Speed up or stabilize inference with advanced samplers and schedulers

Recommended: use ComfyUI-Manager to browse, install, update, and pin versions.

Manual install overview:

1. Navigate to your ComfyUI/custom_nodes directory
2. Clone the node repo into custom_nodes
3. Install Python requirements listed by the node pack
4. Restart ComfyUI to load nodes

Example (generic):

cd ComfyUI/custom_nodes
git clone <repo-url>
cd <repo-folder>
pip install -r requirements.txt

Tip: keep a separate virtual environment and pin working commits before updating large graphs.

These widely used packs cover most anime tasks. Search on GitHub by name to install.

• ComfyUI-Manager: one-click install/update of custom nodes
• WAS Node Suite for ComfyUI: utilities, filters, prompts, image ops
• ComfyUI-Impact-Pack: advanced samplers, conditioning, quality-of-life nodes
• comfyui_controlnet_aux: lineart, canny, depth, normal, and other preprocessors
• ComfyUI_IPAdapter_plus (or similar): image prompt guidance for character/style
• ComfyUI-AnimateDiff-Evolved: animation nodes for motion in anime scenes
• Real-ESRGAN nodes: upscale, especially anime-focused models (e.g., anime6B)
• Optional: LoRA/embedding helpers included in many packs for easy stacking

Goal: crisp lines, flat shading, minimal artifacts.

Pipeline outline:

• Load Checkpoint (anime-ready model) → CLIP Text Encode (positive/negative)
• LoRA Loader (optional, 0.5–0.8 weight for characters)
• KSampler (Euler a or DPM++ 2M Karras; 20–28 steps; CFG 4.5–7)
• VAE Decode → Save Image

High-res pass (optional):

• Upscale (Real-ESRGAN anime model, 1.5–2x)
• KSampler refine (lower denoise 0.15–0.35) → VAE Decode → Save Image

Prompting tips:

• Add tags like “clean lineart, flat shading, cel shaded, anime style”
• Use negative prompts for “blur, noise, artifacts, extra fingers, text, watermark”

ControlNet helps lock pose/composition based on edges.

• Preprocess: Image → ControlNet AUX (Lineart/Canny) → ControlNet Condition
• Main path: Text Encode → KSampler → VAE Decode
• Apply ControlNet: connect condition to KSampler; start weight 0.6–0.9

Tips:

• Use Lineart for clean manga/anime outlines; Canny for general edges
• Adjust guidance strength downward for softer adherence

Combine a character LoRA with IP-Adapter for stronger identity lock:

• IP-Adapter: plug an image reference into the conditioning path; weight 0.5–0.9
• LoRA Loader: start at 0.6–0.8; raise if features drift
• Use seed locking and fixed sampler to stabilize features across shots

For repeating outfits/hair color, add those tokens to the positive prompt and keep CFG ≤ 7.

For manga panels or low-res anime frames:

• Load Image → VAE Encode (or use img2img node) → KSampler
• Denoise strength 0.25–0.5 for cleanup; lower for preserving line art
• Upscale with Real-ESRGAN anime model (1.5–4x) before a light refine pass

Artifacts fix:

• Use WAS/Impact image filters (sharpen lite, mild denoise) before the final save

Add motion without full video models:

• Load AnimateDiff nodes → connect motion module to KSampler chain
• Use short sequences (8–24 frames) and loopable seeds
• Lower CFG (3.5–6) and denoise (0.2–0.35) to reduce flicker

For consistent characters: keep the same LoRA/IP-Adapter refs across frames.

Settings that help 6–12GB GPUs:

• Use fp16 models/VAEs when available
• Reduce resolution (e.g., 768→640) and upscale later
• Batch size 1; keep only 1–2 ControlNets active at once
• Prefer KSampler schedulers like DPM++ 2M Karras for speed/quality balance
• Enable memory-efficient attention (if supported by your build)
• Cache models and avoid frequent checkpoint swaps mid-session

Missing node/class: the pack didn’t load. Reinstall requirements and restart.

Torch/CUDA mismatch: install a Torch build matching your CUDA; verify with a simple base graph.

API/version breaks: pin node pack commits; check release notes after ComfyUI updates.

Model path errors: confirm model folder mapping in node settings; avoid spaces in paths.

ControlNet/Adapter no effect: raise guidance weight slowly; verify correct connection into the KSampler conditioning input.

Best practices:

• Keep a staging environment to test new nodes
• Pin working commits for production graphs
• Read each repository’s license; confirm usage for commercial work
• Export and back up your graphs and custom settings regularly