README.md

metadata

tags:
  - quantized
  - safetensors
  - diffusers
  - mecha
  - weight-only
library_name: diffusers

Wan2.2-A14B-Diffusers

Quantised media packages of Wan-AI/Wan2.2-A14B.

MECHA Forge

MECHA — Model Engineering & Contextual Hearth Assembly

These media packages have been forged and optimized in the MECHA Forge.

Available models

• Wan2.2-I2V-A14B (image-to-video)
• Wan2.2-T2V-A14B (text-to-video)

Description

Wan 2.2 A14B: high/low noise MoE video model. 40 steps, CFG 5.0. Text encoder (umt5-xxl) shared.

Available quantizations

Quant	Denoiser	Text encoder	VAE
FP8	FP8	FP8	FP8
INT8	INT8	INT8	INT8
FP4	FP4	FP4	FP8
INT4	INT4	INT4	INT8

The VAE is shared between formats: FP4 reuses the FP8 file, INT4 reuses the INT8 file. Text encoder (umt5-xxl) is shared across I2V and T2V.

Components

In addition to the denoiser, text encoder, and VAE, the package includes:

• transformer (high_noise + low_noise for I2V/T2V)
• connectors, scheduler, tokenizer, processor

Recommended settings

From the source model card (adjust to taste):

Parameter	Value
Steps	40
CFG / Guidance scale	5.0
Resolution	480P or 720P
Notes	High/low noise MoE; text encoder (umt5-xxl) shared

Example prompts

Caught in the moment (candid, slice of life):

masterpiece, best quality, highres, absurdres, very aesthetic, safe, 1man, solo,
candid moment, natural lighting. A rooftop bar in the city at night, skyline
lights, a glass on the table, cool breeze. A 34-year-old man with a strong jaw
and grey eyes. He has a lean build and a relaxed expression. He wears a linen
shirt and chinos, holding a drink, mid-conversation, candid and relaxed.

Fashion photography (magazine cover style):

Fashion video still: a 34-year-old man with a strong jaw, grey eyes, and salt-
and-pepper hair; wearing a tailored navy suit, pocket square; city rooftop,
golden hour, cinematic, 24fps, 8K.

Scenery:

Scenery: a city skyline at night, lights reflecting on water, skyscrapers,
bridges, cinematic wide shot, 8K.