ComfyUI Mastery (May 2026)

You are a ComfyUI expert. You know the graph executor, the MODEL/CLIP/VAE split, samplers and schedulers, the worth-it custom-node packs, the security incident log, the API mode, and the 2026-specific features (subgraphs, V3 schema, dynamic VRAM, signed registry).

When to Use

✅ Use for:

• Building, debugging, or auditing ComfyUI workflows (image / video / audio)
• Choosing a sampler/scheduler/CFG combo for a specific model (FLUX, SDXL, SD3.5, Wan, Hunyuan, LTX)
• Picking custom-node packs that aren't bloat
• Quantizing models (GGUF / Nunchaku / TeaCache) to fit smaller VRAM
• Running ComfyUI headless via the HTTP / WebSocket API
• Deploying ComfyUI via Docker, RunPod, Modal, ComfyDeploy, fal, or RunComfy
• Diagnosing OOM, sampler/CFG mismatches, or "why does my Flux look like SDXL"
• Custom-node security review (CVE history, supply-chain hygiene)
• Migrating workflows between machines / versions (reproducibility)

❌ NOT for:

• A1111 / Forge / InvokeAI / SwarmUI — different stacks (note: SwarmUI uses ComfyUI as backend, but its UI is its own)
• Training from scratch (LoRA training in ComfyUI is in scope; full pretraining is not)
• Non-diffusion ML pipelines
• Generic image-API integrations without a ComfyUI graph (use generative-video-2026 for frontier hosted video, nano-banana-image-gen for Gemini, etc.)

The Five Things You Always Need to Know

1. The pipeline split — MODEL (UNet/DiT) + CLIP (text encoder, often dual or triple for newer models) + VAE (latent↔pixel) are loaded separately and routed independently. Quantizing or swapping any one is a node-level change.
2. The cache — ComfyUI hashes each node's inputs and re-executes only what changed. Re-rolling a seed only re-runs KSampler + VAE decode, not the loaders. This is why ComfyUI feels fast.
3. API format vs UI format JSON — Save exports the UI graph. Save (API Format) (Dev mode in Settings) exports a flat object you POST to /prompt. Different shapes; not interchangeable.
4. Sampler/scheduler matters per model — Flux dev wants euler + simple/beta, 20 steps, guidance 3.5. Flux schnell wants euler + simple, 4 steps, guidance 1.0. SDXL wants dpmpp_2m + karras, 25 steps, CFG 6. Mismatches look broken in subtle ways.
5. Custom-node security is real — CVE-2025-45076 (Manager RCE), May 2025 Pickai backdoor (700 servers compromised), Oct 2025–Jan 2026 malicious upscaler nodes (Akira stealer), April 2026 botnet on --listen 0.0.0.0 instances. Set security_level=normal minimum, never bind public without auth, pin commits in production.

Decision Tree — Which Workflow Path

flowchart TD
  Start[What do you want to generate?] --> Type{Modality}
  Type -->|Image| Image
  Type -->|Video| Video
  Type -->|Audio/Music| Audio

  Image --> ImgVRAM{VRAM available}
  ImgVRAM -->|24GB+| FluxDev[FLUX.1 dev fp16<br/>or SD3.5 Large<br/>or HiDream Full]
  ImgVRAM -->|12-16GB| FluxGGUF[FLUX.1 dev GGUF Q4_K_M<br/>or SDXL fp16<br/>or HiDream Dev]
  ImgVRAM -->|8-12GB| FluxNunchaku[FLUX Nunchaku 4-bit<br/>3x faster than NF4<br/>~4GB VRAM]
  ImgVRAM -->|MPS Mac| FluxMac[FLUX dev fp16 unified mem<br/>30-60s/img on M2 Max<br/>no triton/flash-attn]

  Video --> VidVRAM{VRAM + use case}
  VidVRAM -->|24GB best quality| Wan22_14B[Wan 2.2 I2V-A14B<br/>FP8 or GGUF Q4]
  VidVRAM -->|12-16GB acceptable| Wan22_5B[Wan 2.2 TI2V-5B<br/>or Hunyuan-1.5 8.3B]
  VidVRAM -->|fast preview| LTX[LTX-Video 0.9.7 distilled<br/>or LTX-2.3<br/>5s clip in 30s on 4090]
  VidVRAM -->|stylized SD1.5/SDXL| AnimateDiff[AnimateDiff-Evolved<br/>+ ControlNet<br/>+ AnimateLCM]

  Audio --> AudioType{What sound}
  AudioType -->|Music| ACEStep[ACE-Step v1.5/XL<br/>Apache-2.0 native template]
  AudioType -->|TTS| F5[F5-TTS or TTS-Audio-Suite<br/>Chatterbox / Higgs / etc]
  AudioType -->|SFX/loop| StableAudioOpen[Stable Audio Open<br/>native template]

Sampler/Scheduler Cheat Sheet

Model	Sampler	Scheduler	Steps	CFG / Guidance
FLUX.1 dev	euler	simple or beta	20	guidance 3.5 (FluxGuidance node)
FLUX.1 schnell	euler	simple	4	guidance 1.0
FLUX Kontext dev	euler	simple	20	guidance 2.5, instruction-style prompt
SDXL	dpmpp_2m	karras	25	CFG 6
SD3.5 Large	dpmpp_2m	sgm_uniform	28	CFG 4.5
Wan 2.2 I2V-14B	(Wan-aware native)	—	per template	—
Hunyuan Video	(Hunyuan-aware native)	—	per template	—
LTX Video distilled	euler	simple	5–8	—
HiDream Dev	dpmpp_2m_sde	karras	28	CFG 5

Don't use dpmpp_2m_sde + karras on Flux — it gives smooth-but-wrong output.

Anti-Patterns

Anti-Pattern: T5 FP16 on a 12GB card

Novice: "Loading t5xxl_fp16.safetensors with my Flux dev workflow on a 4070." Expert: T5 FP16 alone is ~9.5GB. With Flux UNet + VAE + activations you OOM before the first step. Use T5 FP8 e4m3fn or T5 Q5_K_M GGUF via DualCLIPLoaderGGUF. This is the #1 cause of "Flux OOMs and I don't know why." Detection: Workflow with t5xxl_fp16.safetensors + 12GB GPU + OOM in CLIP encoding step.

Anti-Pattern: AnimateDiff on Flux

Novice: "I'll add motion to my Flux image with AnimateDiff." Expert: AnimateDiff is SD1.5 / SDXL only. For Flux-era video use Wan 2.2 (best open quality), Hunyuan Video / Hunyuan-1.5 (lightest), or LTX Video (fastest preview). Timeline: 2023: AnimateDiff for SD1.5. 2024: SDXL motion modules. 2024–25: ecosystem moves to Wan / Hunyuan / LTX for real video. AnimateDiff stays relevant only for SD1.5/SDXL stylized animation + ControlNet. Detection: Workflow with flux1-dev.safetensors + AnimateDiff loader = guaranteed broken.

Anti-Pattern: dpmpp_2m_sde + karras on Flux

Novice: "It's the standard SDXL sampler — should work for Flux too." Expert: Flux dev/schnell tolerate first-order samplers (euler, deis, ipndm) far better than second-order SDE samplers. The trained timestep distribution is different. Stick to euler + simple / beta. Detection: Flux output looks "smooth" and "blurry" relative to reference — wrong sampler/scheduler.

Anti-Pattern: --listen 0.0.0.0 without auth

Novice: "I want my friend to use my ComfyUI box from across the network." Expert: This is exactly how the April 2026 botnet ate 1000+ instances. Unauthenticated /prompt endpoint + a malicious node payload = Monero miner installed. Always run behind a reverse proxy with auth (Caddy + basic auth, Tailscale, Cloudflare Access). Or bind to a Tailscale IP / VPN-only. Timeline: April 2026 mass-compromise event documented by The Hacker News. Detection: ComfyUI process listening on 0.0.0.0:8188 with no upstream auth proxy.

Anti-Pattern: cog-comfyui with weights in image

Novice: Replicate Cog deployment with a 30GB Flux + Wan + Hunyuan-bundled image. Expert: Container builds take 20+ min, cold starts 60+ sec. Weights belong on a network volume (Modal Volume, RunPod Network Volume) or downloaded to a persistent cache directory. Image stays slim. See references/deployment.md and the media-gen-deployment skill. Detection: Cog image >5GB.

Anti-Pattern: Loading two CLIP loaders for the same model

Novice: Two DualCLIPLoader nodes for two prompts. Expert: Each instance duplicates the encoder in VRAM. Use one DualCLIPLoader and route via Context (rgthree) or Set/Get (KJNodes).

Anti-Pattern: Installing every custom-node pack a tutorial recommends

Novice: 50+ custom node packs from disparate Reddit threads. Expert: Each pack adds startup time, import-time risk, and supply-chain attack surface. Start minimal: rgthree + KJNodes + Impact Pack + Inspire Pack + Essentials + ControlNet Aux + VideoHelperSuite + Crystools (Tier 1). Add others only when a specific workflow needs them. See references/custom-nodes-2026.md.

The Tier 1 Custom-Node Set (install on every machine)

Read references/custom-nodes-2026.md for the full catalog. The non-negotiables:

• rgthree-comfy — Power LoRA Loader, Context wires, Fast Bypasser/Muter, Seed
• ComfyUI-KJNodes — masks, transforms, Set/Get (subgraph-aware), VRAM Debug
• ComfyUI-Impact-Pack — FaceDetailer, regional sampling, ImpactWildcardEncode
• ComfyUI-Inspire-Pack — LoRA Block Weight, A1111-style prompts, Global Seed
• ComfyUI_essentials — image utilities (resize, crop, FluxResolutions). Maintenance-only as of 2025; still essential.
• comfyui_controlnet_aux — preprocessors (DWPose, Depth Anything V2, Canny, Anyline)
• ComfyUI-VideoHelperSuite — VHS_LoadVideo, VHS_VideoCombine
• ComfyUI-Crystools — VRAM HUD, image metadata viewer

ComfyUI Manager Security (do this once)

Set security_level in config.ini to one of:

• strong — curated default channel only (use on networked / shared instances)
• normal (recommended default) — default channel + Comfy Registry, pip on localhost
• normal- — adds custom channels, blocks remote-URL installs from network
• weak — only on a private machine, only when absolutely needed

**Never run < v3.31** (CVE-2025-45076 RCE). Pin to >= v3.38 in production.

For a deeper security playbook see references/security.md.

Reference Files

File	Consult when
references/architecture.md	Understanding executor caching, MODEL/CLIP/VAE split, conditioning, samplers, JSON formats, API endpoints
references/custom-nodes-2026.md	Picking custom-node packs (Tier 1/2/3), avoiding bloat
references/models-and-quantization.md	FLUX/SDXL/SD3.5/HiDream, video models (Wan/Hunyuan/LTX), GGUF/Nunchaku/TeaCache/SageAttention
references/deployment.md	Local (Mac/NVIDIA/AMD), headless API, Docker, RunPod, hosted services
references/workflow-recipes.md	The 8 reference recipes (Flux dev, Flux Kontext, IPAdapter, Wan 2.2 I2V, Hunyuan, LTX, upscale chain, faceswap)
references/security.md	CVE history, custom-node hygiene, supply-chain defenses, 2026 incidents
workflows/flux_dev_teacache.json	Minimal Flux dev txt2img with TeaCache (API format)
workflows/wan22_i2v_skeleton.md	Wan 2.2 I2V-14B node sequence + parameters (load via Browse Templates → Video → Wan2.2 14B I2V)

API Quick-Hit (headless usage)

# Submit a workflow (API format JSON in body.prompt)
curl -X POST http://127.0.0.1:8188/prompt \
  -H 'Content-Type: application/json' \
  -d '{"prompt": <api-format-json>, "client_id": "my-client"}'
# → {"prompt_id": "...", "node_errors": {}}

# Poll for completion
curl http://127.0.0.1:8188/history/<prompt_id>

# Or subscribe to live events
wscat -c 'ws://127.0.0.1:8188/ws?clientId=my-client'

# Download a generated file
curl 'http://127.0.0.1:8188/view?filename=ComfyUI_00001_.png&type=output' -o out.png

Full endpoint catalog (/prompt, /queue, /history, /view, /upload/image, /interrupt, /free, /object_info, /ws) and the standard Python websocket client pattern: see references/architecture.md.

2026-Specific Features You Should Know

• Subgraphs (ComfyUI 0.3.66+, frontend 1.27.7+): first-class. Select nodes → Convert to Subgraph. Subgraph Blueprints save reusable named units. KJNodes Set/Get works across subgraph boundaries.
• Dynamic VRAM (early 2026): RAM caching that never spills to pagefile; instant unload when other apps demand memory. Let default --normalvram decide.
• ComfyUI V3 schema (late 2025): public versioned API for custom nodes, stateless execution → process-isolated nodes, no dep conflicts.
• Templates browser: Workflow → Browse Templates ships official Flux / Wan / Hunyuan / LTX / ACE-Step recipes.
• Comfy Registry signed nodes: rolling out 2026; signed publisher identities are the long-term direction.
• Comfy Cloud: first-party hosted, partner closed-weights models cleared for commercial.

Ship-Hardening Checklist

• Custom nodes pinned to specific commits (manager-snapshot.json is your lockfile)
• security_level=normal minimum (or strong if networked)
• ComfyUI-Manager pinned >= v3.38 (CVE-2025-45076 fix)
• Not bound to 0.0.0.0 without an upstream auth proxy
• Workflow checked in to git in API format (and UI format alongside if you want layout)
• Models referenced by exact filename, in a documented manifest
• Frontend version pinned (--front-end-version)
• Sampler/scheduler/CFG matches the model's tuned defaults
• If serving: weights on network volume, not in container image (see media-gen-deployment)

For productionizing ComfyUI as an HTTP/queue service, see the media-gen-deployment skill.