GeoFlow

GeoFlow is a PEFT LoRA adapter for Wan2.1-T2V-1.3B-Diffusers. The adapter is trained for text-to-video generation with an emphasis on improved geometric and temporal consistency.

This repository contains the released GeoFlow model weights.

Release Status

• Model weights
• Inference code
• Evaluation code
• Reward model code

Model Details

Field	Value
Model type	LoRA adapter
Base model	wan-ai/Wan2.1-T2V-1.3B-Diffusers
Framework	PEFT / Diffusers
Checkpoint format	safetensors
Adapter rank	r = 32
LoRA alpha	64
LoRA dropout	0.0

Adapter Targets

GeoFlow adapts attention projection modules in the Wan transformer:

to_q, to_k, to_v, to_out.0,
add_q_proj, add_k_proj, add_v_proj, to_add_out

Repository Files

File	Description
adapter_model.safetensors	GeoFlow LoRA weights
adapter_config.json	PEFT adapter configuration
README.md	Model card

Intended Use

GeoFlow is intended for research use in text-to-video generation. It should be loaded as a LoRA adapter on top of the Wan2.1-T2V-1.3B Diffusers checkpoint.

The released adapter can be used to sample videos from text prompts, compare geometry-aware video generation methods, and support reproducible evaluation once the eval and reward model code are released.

Inference

Reference inference code is available here:

https://github.com/geometryflow/GeoFlow

The adapter can also be downloaded directly:

huggingface-cli download ackermannj/geoflow --local-dir checkpoints/geoflow

Limitations

GeoFlow inherits limitations from the Wan2.1 base model. Generated videos may still contain artifacts, object distortions, temporal inconsistencies, incorrect motion, or prompt-following failures. Outputs should be reviewed before use in downstream applications.

Citation

Citation information will be added with the accompanying project release.