I set up the Banostasis concept LoRA on WAN 2.2 T2V 14B and ran it in ComfyUI. It is a high noise LoRA meant for fast, rhythmic scene cuts. If you follow the exact prompt pattern the author suggests, it locks into a very specific look and motion.
What this LoRA does
- Adds quick scene transitions with bold lighting and clean camera moves
- Works only with a high noise sampler in your pipeline
- Responds best when you include two fixed phrases in the prompt
- Built for the Wan2.2 T2V A14B base model family
- Inference tested in ComfyUI by the author
These points are called out on the model card.
Download and files
Grab the safetensors from the Files tab. The current file on the page is:
246839-wan22_14B-high-banostasis_concept-e459.safetensors at 307 MB.
Model page for reference. Hugging Face
The exact prompt pattern
This LoRA is picky about wording. The author asks you to include these stems.
- Banostasis concept
- Sequence of quick scene transitions of
- The sequence of changes includes the transitions
Use them up front and your results stabilize. Skip them and the rhythm drifts
Copy-paste starter
Banostasis concept, a sequence of quick scene transitions of [subject],
the sequence of changes includes the transitions between [location 1], [location 2], [location 3],
shot with [camera move] and [lighting], stylized in [visual style].
ComfyUI setup
- Load Wan2.2 T2V A14B as base
- Load the Banostasis LoRA at strength 0.8 to 1.0
- Pick a high noise sampler
- Paste the prompt with the fixed stems above
- The author uses ComfyUI for inference and recommends that strength range. The sampler should be high noise only.
Working examples you can try
The page shows several prompt ideas that match the style. I rewrote them slightly to keep them concise while keeping the structure.
Musical instruments dissolving into sound waves and glowing orbs, cutting between a concert hall, a subway tunnel, and a desert plain, with fast push-ins and neon abstract lighting.
A monk kneeling in prayer and holding still while the world changes, moving through a temple hall, a mountain cave, a crowded street, and a digital gridscape, with circling dolly shots under firelight and neon glow.
Painted masks peeling to reveal new faces, switching from a festival to an empty theater stage to a dim shrine, with close telephoto shots and side light.
A woman in a red dress, eyes closed, standing still as the scene jumps from cathedral to subway platform to snowy cliff to neon skyline, with lens push-ins and soft side light.
Balloons turning into jellyfish underwater and then into paper lanterns floating at night, moving between city festivals, deep sea, and mountain temples, with sweeping wides under moonlight.
Training notes
- Trained only on videos
- Dataset: 33 clips at 480×270 with 25, 33, 65, 81 frames per clip
- Steps: 15147
- LR: 5e-5
- Optimizer: AdamW
- Rank: 32
- Batch size: 1
- Grad accumulation: 1
- min_t 0.875, max_t 1
- Trained with the diffusion-pipe repo
- These details are listed on the model card. The repo is public and used for WAN 2.2 training.
