Created by HearmemanAI https://www.hearmemanai.com

Troubleshooting guide in case you encounter any errors:

click here

v12 — Major Refactor (May 2026)

⚠️ Breaking change for existing users. The old per-resolution / per-feature flags are gone. If you're updating an existing pod template, delete the removed env vars below and add the new ones — otherwise nothing will download.

⚠️ CUDA 12.8 required. Before deploying, open Additional Filters and select CUDA 12.8. Fallback (older CUDA): https://runpod.io/console/deploy?template=0b7tk92912&ref=uyjfcrgy

---

What's new in v12

• Workflow-driven downloads. The pod now reads the workflows you choose and downloads only the models those workflows actually reference. No more redundant multi-gigabyte downloads.
• Faster HF transfers. Hugging Face's hf_xet accelerator is enabled image-wide (HF_XET_HIGH_PERFORMANCE=1), typically 1.5–3× faster than aria2 on HF.
• Live download manager. A 3-way pool with an append-only status snapshot every 10s (active %, speed, ETA; queued list; done count). No more interleaved aria2 spam.
• Faster boot. comfy-aimdo / comfy-kitchen now install in the background instead of blocking startup; redundant apt/pip steps removed.
• Slimmer workflow set. Obsolete Wan-Fun-SDXL helper, legacy VACE preview, and the unused Video Extend workflow have been removed.

---

🧪 Experimental: CUDA 13 + NVFP4 (-cuda13 tag)

Every release also ships an experimental CUDA 13 image. To use it, append -cuda13 to the image tag (e.g. …/comfyui-wan-template:v12-cuda13 instead of :v12), or deploy from the experimental template.

• What you get: PyTorch cu130, so NVFP4 quants load natively on Blackwell GPUs (RTX 50xx / sm_120). On other GPUs NVFP4 degrades to fp16/fp8 — no speedup, but it still runs.
• SageAttention ships as a prebuilt cu130 wheel and is used automatically when your GPU supports it; otherwise it's built from source on first boot (same as the standard image).
• Deploy with CUDA 13.0+ selected in Additional Filters.
• The standard :vN image (CUDA 12.8) is unchanged — stay on it if you don't specifically need NVFP4.

---

⚙️ Environment Variables

Set the flags you want to true (lowercase, as strings). Multiple may be enabled at once — shared models are deduplicated automatically.

Download flags (new in v12)

Variable	Downloads + provisions
download_wan21	Wan 2.1 (incl. VACE) + Infinite Talk workflows + their models
download_wan22	Wan 2.2 + SVI Video Extension workflows + their models
download_wan_animate	Wan Animate workflows + models
download_steady_dancer	Steady Dancer workflow + models

Only the workflows for enabled flags are copied into your ComfyUI workflows folder — you'll see exactly what your models can run.

Removed in v12 (delete these from your template if you're upgrading)

download_480p_native_models, download_720p_native_models, download_wan_fun_and_sdxl_helper, download_vace, download_vace_debug, download_wan_animate_only, debug_models

CivitAI (unchanged)

Variable	Description
civitai_token	Your CivitAI token (auto-download LoRAs and Checkpoints)
LORAS_IDS_TO_DOWNLOAD	Comma-separated CivitAI LoRA version IDs
CHECKPOINT_IDS_TO_DOWNLOAD	Comma-separated CivitAI Checkpoint version IDs

👉 CivitAI Downloader README

---

🚀 Deploying

1. Set your env vars (at least one download_* flag set to "true").
2. Click Deploy.
3. Wait for setup (5–20 minutes depending on flags + network).
4. Future deployments are faster on a persistent network volume.

🌐 Accessing ComfyUI

1. Click Connect
2. Open port 8188

📓 Accessing JupyterLab

1. Click Connect
2. Open port 8888

---

🛠️ Troubleshooting

Fixing Missing Custom Nodes

If you see IMPORT FAILED in ComfyUI:

1. Open the Manager menu
2. Click Install missing custom nodes
3. Click Try fix

User-supplied LoRAs

If a workflow references a LoRA that isn't bundled (your personal training LoRAs, NSFW pack, etc.), the boot log will list them as "user-supplied". Drop them into /workspace/ComfyUI/models/loras/ manually or via LORAS_IDS_TO_DOWNLOAD.

ℹ️ For my other templates: Click HERE