Wan 2.2 Remix + PromptRelay in ComfyUI: GGUF workflow guide

This guide documents a local ComfyUI workflow for Wan 2.2 Remix, using the GGUF-capable shared workflow highlighted in the video. The practical goal is a repeatable image-to-video or text-to-video setup that can generate short cinematic clips with realistic motion, scene consistency, and segmented prompt control.

Keep your model files, outputs, and prompts in a private workspace. If a workflow or model card includes its own license, prohibited uses, or distribution rules, follow those terms. For public or professional demonstrations, use safe, non-explicit prompts and label AI-generated media clearly.

• Use fictional adult subjects only.
• Avoid celebrity, coworker, student, or private-person likenesses unless you have explicit permission.
• Do not upload sensitive source images into public services.
• Review outputs before sharing; discard anything unsafe or unintended.

The tutorial describes the graph in practical zones. Think of it as a left-to-right production line.

1. Download the shared workflow. Use the Google Drive workflow link from the video description, then save the workflow JSON somewhere easy to find.
2. Choose the Remix model format. The video says Remix variants are available as safetensors and GGUF. GGUF is often the practical choice for constrained systems; safetensors may be preferable when you have enough VRAM and a compatible loader.
3. Select version 3 for both model stages. For the workflow shown, use the version 3 variant for both the high-noise and low-noise loaders. The transcript notes that FP16 is available for some variants, but v3 FP16 was not available at the time of the video.
4. Download both high and low models. The graph expects a high-noise model and a low-noise model. If only one is connected, the workflow will fail or produce poor results.
5. Download the text encoder / CLIP models. The video uses adult-content-capable text encoder links and mentions GGUF variants as direct downloads.
6. Place text encoders in the ComfyUI model folder. Use your ComfyUI text encoder folder, commonly ComfyUI/models/text_encoders/. If your custom nodes expect a different folder, follow the node documentation.

ComfyUI/
  models/
    diffusion_models/   # common location for video/diffusion model files
    text_encoders/      # place text encoder / CLIP files here when required
    loras/              # optional acceleration/flavor LoRAs
    vae/                # VAEs if the workflow requires them

Folder names can vary by custom node. When a loader dropdown does not show your model, check the node’s expected model directory and restart ComfyUI or refresh model lists.

The shared workflow includes color-coded loader groups. The video shows both regular split-file loader nodes and the creator’s preferred loader nodes with more optimization controls.

1. Open the workflow JSON in ComfyUI. If nodes are red or missing, install the missing custom nodes first, restart, and reload the workflow.
2. Set the high-noise loader. Choose the high-noise Wan 2.2 Remix v3 model in the loader intended for the high flow.
3. Set the low-noise loader. Choose the matching low-noise Wan 2.2 Remix v3 model in the low-flow loader.
4. Reconnect loader outputs if needed. The video demonstrates rewiring by holding Shift and dragging the connection to the target input.
5. Handle optional LoRAs. If you are not using acceleration or flavor LoRAs, toggle them off or bypass those loader nodes entirely.
6. Choose text encoder handling. For GGUF CLIP/text encoder, the creator offloads to a secondary GPU. If you do not have a second GPU, disable or delete the set-device node and let the workflow use your normal device path.

PromptRelay lets you describe a clip as timed segments instead of one long prompt. The workflow can switch between regular prompt input and PromptRelay input.

1. Open ComfyUI Manager.
2. Choose Install via Git URL.
3. Paste https://github.com/kijai/ComfyUI-PromptRelay.
4. Confirm installation and restart ComfyUI.
5. In the workflow’s prompt switch, set the control to use PromptRelay. The video says false selects the PromptRelay node and true selects the regular prompt node.

wide cinematic shot, adult character enters a neon hallway |
character turns toward the camera, coat moving naturally |
camera pushes in, lights flicker, character exits frame

The settings group controls total steps, split steps, and video length. The resize/preview node controls output size for the image-to-video path.

1. Load a test image for I2V. Use the linked test image or your own safe, appropriate source image.
2. Check the preview after resizing. Make sure framing still looks correct after the resize node.
3. Set conservative dimensions. Avoid large resolutions until the workflow runs cleanly.
4. Set total and split steps. Use the workflow defaults first. Later, adjust quality/speed once you understand runtime and VRAM.
5. Queue the prompt and watch the console. Missing model, missing node, and CUDA out-of-memory errors are easiest to diagnose from the console output.

Wan-style video workflows tend to respond best to visual, camera-aware prompts. Segment prompts should describe motion changes, not just static appearance.

Global: cinematic realistic lighting, stable anatomy, consistent outfit, smooth natural motion, detailed environment, safe fictional adult subject

Segments:
subject stands in a softly lit room, camera begins a slow push-in |
subject turns and walks toward the window, fabric moves naturally |
subject pauses, looks outside, warm light crosses the face |
camera settles, subject relaxes, motion resolves cleanly

• The Render Den — Wan2.2 Remix NSFW with PromptRelay node. Primary tutorial source.
• Workflow Google Drive link from the video description.
• KJ's ComfyUI PromptRelay. Prompt segmentation custom node.
• Test image Google Drive link from the video description.
• ComfyUI GitHub repository. Core node-graph interface.
• ComfyUI Manager. Custom node installation helper.

This guide summarizes procedural points from the transcript and description. Model repository links and model license terms may change; always read the model page or repository before downloading or generating.