FollowCam Redmond WAN2 T2V 14B
Wan 2.2 text-to-video LoRA for immersive follow-camera action tracking shots across environments like storms, caves, and salt flats.
On this page
FollowCam Redmond WAN2 T2V 14B
TL;DR
- Trains Wan 2.2 T2V to produce continuous follow-camera motion rather than static scenes.
- Trigger word:
FollowCam; prompt the shot subject and environment explicitly. - Best starting point in ComfyUI is a base Wan model with LoRA strength around 0.8–1.0.
- Output is video-first, so frame selection and motion wording matter more than detail-only prompts.
What the model does
FollowCam Redmond is built for a narrow but visually strong behavior: immersive tracking shots in which the camera follows a subject through a space over time. Instead of asking for a single composed frame, you describe motion and environment, and the LoRA biases the generation toward sustained forward or matching movement.
Sample outputs show a rider on a motorcycle crossing bright salt flats under a wide sunlit sky, a diver swimming through a narrow underwater limestone passage, and a worker climbing a steel-frame building skeleton. Those are consistent with the author’s claimed use cases and show the model prioritizing camera motion and spatial continuity over静止 detail.
How it is trained and what to expect
This is a diffusion LoRA for the Wan 2.2 T2V-A14B pipeline, distributed as a single safetensors file for ComfyUI. The HuggingFace card lists example prompts in the form FollowCam. A continuous ... shot following ..., each emphasizing a specific perspective, movement style, and environment. That structure is useful because the model appears to respond to both the FollowCam trigger and the motion description rather than either alone.
From inspected samples, the model handles different scales well: open outdoor motion, confined interior motion, and underwater motion. The look is action-cinematic rather than photorealistic documentary, so expect stylized lighting and motion blur consistent with first-person or action-camera footage.
Recommended usage
- Base model: Wan-AI/Wan2.2-T2V-A14B.
- Workflow: load the base model, apply the LoRA, then run KSampler as usual.
- Starting LoRA strength: 0.8–1.0.
- Prompting pattern:
FollowCam. A continuous [shot type] following [subject] through [environment]. - Subject and environment specificity matter more than long style lists.
Common errors and fixes
Error | Cause | Fix | Source
|---|---|---|---
Static-looking shot | Weak motion wording or omitted trigger | Lead prompt with FollowCam. A continuous ... shot | HF README Warped subject tracking | Very high LoRA strength with noisy base | Lower strength_model slightly and increase sampling steps | Civitai Watery underwater artifacts | Underwater motion is hard for T2V | Add underwater, cave, and lighting cues like headlamp | HF examples
FAQ
Does this work in A1111? The card targets ComfyUI and lists Wan-based generation. Use the same base Wan model and apply the LoRA with your preferred Wan loader if available.
Can it do non-action scenes? It is trained for follow-motion. Static or slow dialogue scenes are not the intended use case.
Is the video consistent? Follow-camera consistency is the core behavior. Shorter clips with explicit environmental constraints tend to stay on subject longer.
Sources
- https://huggingface.co/artificialguybr/FollowCam-Redmond-WAN2-T2V-14B
- https://civitai.com/models/2068735