Woovee/Wan2.2-14B-Studio

Wan 2.2 A14B Studio package

Public, attributed package of public weights for T2V and I2V. Base experts use Comfy-Org's FP16 repackaging of Wan-AI weights and the original configs, avoiding the original FP32 download size. No Studio source, keys or generated outputs are included.

• t2v/base: official Wan-AI T2V experts. Use the bundled LightX2V 1217 acceleration pair, 4 DMD steps, CFG disabled, boundary step 2.
• i2v/fast: native LightX2V FP8 conversion of April 12 2026 distilled experts. Acceleration is baked in; do not stack a speed LoRA.
• i2v/base: official I2V experts for alternative acceleration presets.
• shared: original UMT5 text encoder/tokenizer and Wan 2.1 VAE needed by both A14B modes.
• loras/t2v/acceleration: required high/low speed pair, applied automatically at strength 1.
• loras/i2v/slow-motion-fix: alternative 4-step high/low pair by rzgar, on the base I2V experts, strength 1. Experimental until tested.
• loras/i2v/weak-id: rzgar's small patch-embedding adapters. Suggested starting strength 0.3; more prompt influence trades away reference identity. Experimental until tested.

See catalog.json for Studio presets and reviewed adapter routing. Files placed in arbitrary folders are not automatically compatible. Wan 5B adapters cannot be used here.

Use model/block CPU offload on 48 GB Ada GPUs, with FP8 quantization where configured. Weight file sizes are not peak VRAM measurements. Native Wan video uses 16 fps; dimensions must follow the selected runtime's constraints. Minimal tests are not proof of visual quality at every resolution.

Each original repository, pinned revision, filename, byte size and SHA-256 is in provenance.json. Model cards are retained under attribution/. Model authors: Wan-AI/Alibaba; acceleration/runtime: LightX2V/ModelTC; April FP8 conversion: gufisha; motion and reference-control adapters: rzgar. The mirrored upstream packages declare Apache-2.0; owner-added files retain their own original licensing terms. Optional adapters are not a guarantee of better results and have not yet been output-tested in this deployment.

Additional selectable pairs

loras/i2v/acceleration contains the original LightX2V 1022 high/low pair, used with base I2V experts as a separate four-step preset. Never stack it with the fast distilled checkpoint or motion-correction acceleration.

The owner added DR34ML4Y_I2V_14B_HIGH/LOW_V2 and NSFW-22-H/L-e8 files. Their 40-block, 5120-channel low-rank tensor shapes were inspected. They are optional paired adapters with user-controlled strength; shape compatibility is not output-quality verification or a licensing endorsement. Their files were preserved.