Wan 2.1 guide: checkpoints, setup & current alternatives

Status checked September 13, 2026. This guide does not declare Wan 2.1 the best model or promise a universal VRAM floor, render time or quality result. Those depend on the checkpoint, resolution, frame count, dtype, offloading, attention implementation and hardware.

Choose the correct Wan 2.1 checkpoint

• T2V-1.3B: the smaller text-to-video route. The repository says it supports 480p and can run on consumer GPUs with 8.19 GB VRAM; treat that as the project’s documented configuration, not a guarantee for every environment.

• T2V-14B: the larger text-to-video checkpoint for 480p or 720p generation.

• I2V-14B-480P or I2V-14B-720P: image-to-video checkpoints. Match the checkpoint to the intended output resolution.

• FLF2V-14B-720P: first-and-last-frame guidance at 720p.

• VACE-1.3B or VACE-14B: video creation and editing checkpoints. Follow their task-specific examples rather than assuming the T2V command applies.

Install and run the official repository

The official dependency file specifies PyTorch 2.4 or newer. Use a Python environment compatible with those dependencies and follow the repository’s CUDA and FlashAttention notes for your machine. The minimal repository setup is:

git clone https://github.com/Wan-Video/Wan2.1.git
cd Wan2.1
pip install -r requirements.txt

Download only the checkpoint you intend to run from the official Hugging Face or ModelScope links in the repository. Then copy the matching command from the repository. Do not mix a 1.3B command, a 14B checkpoint and unrelated community settings.

A matching 1.3B command and memory recovery

After downloading the T2V-1.3B checkpoint into the named directory, this documentation-derived example keeps the task, checkpoint and 480p size aligned. The prompt is an illustrative starting brief, not a tested output:

python generate.py --task t2v-1.3B --size "832*480" --ckpt_dir ./Wan2.1-T2V-1.3B --offload_model True --t5_cpu --sample_shift 8 --sample_guide_scale 6 --base_seed 42 --prompt "A ceramic mug on a wooden table. One continuous slow camera push. The mug stays still."

The repository recommends guidance scale 6 and a shift between 8 and 12 for T2V-1.3B. CPU offloading and T5 on CPU are its documented GPU-memory recovery options, not a guarantee that every machine will fit. The command parser otherwise defaults to T2V-14B and 720p; explicitly set the task and size. If changing the frame count, follow its 4n+1 requirement, such as 81 frames. A fixed seed records this run setting, not guaranteed identical output across environments.

A reproducible first test

• Start with the repository’s documented size, frame count and sample command for the selected checkpoint.

• Save the checkpoint name and revision, repository commit, prompt, negative prompt, seed, size, frame count, solver, sampling steps and guidance scale.

• For image-to-video, preserve the exact source image and crop. A changed crop is a changed test.

• Generate at least three attempts before changing one variable. Record runtime, peak memory, failures and accepted outputs.

• Review subject identity, anatomy, object permanence, camera motion, background stability, text and looping artifacts at full size.

Prompt Wan 2.1 without invented guarantees

Describe one shot: subject, action, environment, camera and lighting. For image-to-video, focus on desired motion and what must remain stable. Begin with one subject and one main action. Add complexity only after the baseline works.

A prompt cannot force a checkpoint to preserve every detail. If identity, a product label or composition must be exact, use a source image and inspect every frame. Our video prompt guide explains how to state motion and camera direction across current models.

When to use Wan 2.1 now

• Keep it when an existing workflow, LoRA, benchmark or dependency is pinned to Wan 2.1.

• Use it when the 1.3B route fits constrained local experimentation and you can validate the project’s documented memory configuration on your hardware.

• Choose a current model when starting a new deployment and you do not need 2.1 compatibility.

• Choose a managed API or web product when GPU setup, model downloads, queues and monitoring cost more than the control of self-hosting.

Current alternatives

Wan 2.2 provides newer downloadable checkpoints than 2.1, including T2V and I2V 14B models plus a 5B text-and-image-to-video route documented for 720p at 24 fps. Evaluate those checkpoints when you can change an existing self-hosted pipeline; retain 2.1 when compatibility is the requirement.

Hosted Wan is a separate decision. Alibaba’s current Model Studio catalog lists wan3.0-video and wan3.0-video-prime. These hosted endpoints are distinct from the downloadable 2.1/2.2 repositories. Verify the model in your provider account, its supported task and regional quote before integration; do not treat a higher version number as an installable replacement or as proof of availability in Magic Hour.

LTX 2.5 is a current open audio-video family with official inference tooling. MiniMax H3 publishes downloadable base checkpoints for text/keyframe and multimodal reference audio-video workflows. Compare licenses, hardware, native audio, supported inputs and the complete local-plus-hosted pipeline before selecting either.

The current open-source-friendly video guide compares those routes with managed APIs. Magic Hour’s model catalog lists available hosted models; check the live picker and credit estimate before assuming a particular version is exposed.

Browser alternatives

Use Text to Video when you want to start from a prompt, or Image to Video when a source image defines the subject or composition. Retain the same evaluation record you would use locally: selected model, prompt, source, duration, resolution, attempts and accepted result.