GingerLabs-Wan2.2-T2V-A14B

GingerLabs Wan 2.2 T2V A14B

A GingerLabs-tuned derivative of Wan-AI/Wan2.2-T2V-A14B, packaged as quantized HighNoise and LowNoise diffusion-model pairs for ComfyUI.

This is a quantized-weights-only release. It does not include the VAE, text encoder, or other pipeline assets required for generation.

Content warning: This model may generate mature, explicit, or NSFW material, including unintended outputs. Use it only in lawful, consensual, age-appropriate contexts.

Reference ComfyUI workflow

Download GingerLabs_T2V_Workflow.json

The reference workflow uses ComfyUI-WanVideoWrapper for both GGUF and scaled FP8. It defaults to the balanced Q5_K_M pair, portable sdpa attention, a 720 × 1280 portrait canvas, 33 frames, and 16 fps.

No acceleration LoRA is required.

Required files and nodes

Required text encoder

The reference workflow supports only nsfw_wan_umt5-xxl_bf16.safetensors from NSFW-API/NSFW-Wan-UMT5-XXL. Other UMT5-XXL checkpoints are not supported or validated.

Do not select the sibling nsfw_wan_umt5-xxl_fp8_scaled.safetensors file with this workflow. WanVideoWrapper 1.3.6's LoadWanVideoT5TextEncoder node does not support scaled-FP8 T5 weights.

GingerLabs sampling recipe

Start with these values:

The HighNoise and LowNoise models split one schedule. With 8 total steps, HighNoise runs steps 0–4, then LowNoise continues from 4–end. This is 8 total denoising steps (4 + 4), not 8 steps per model.

Both sampler nodes must use the same total-step count, scheduler, shift, CFG, and seed. The HighNoise end_step must equal the LowNoise start_step. The supplied workflow enforces this with one shared Total Steps control and one shared High → Low Switch control.

Eight steps is the recommended starting point. For a slower refinement pass, try 10 steps with switch 5 or 12 steps with switch 6. Fewer than 8 steps have not been validated.

At CFG 1.0, WanVideoWrapper skips negative conditioning, so the negative prompt has no effect. It becomes active only when CFG is raised above 1.

Download a matching pair

Wan 2.2 T2V A14B uses separate HighNoise and LowNoise models. HighNoise handles the early, high-sigma stage; LowNoise continues the later refinement stage. Download both files for one quantization tier, keep the tiers matched, and do not swap their roles—for example, pair High_Q5_K_M with Low_Q5_K_M.

Sizes use GiB (2^30 bytes). Q5_K_M is the balanced workflow default; use Q8 for a quality-first option or Q4/Q5 when disk and memory are tighter.

Loader settings

In the supplied workflow, use WanVideoModelLoader for both formats:

• GGUF: leave quantization at disabled; WanVideoWrapper loads GGUF directly.
• Scaled FP8: disabled auto-detects the checkpoint's scaled-FP8 marker in WanVideoWrapper 1.3.6+. Selecting fp8_e4m3fn_scaled explicitly is also valid.
• Attention: sdpa is the portable default. SageAttention is an optional speed optimization and requires a compatible installation.

ComfyUI-GGUF remains a supported alternate GGUF workflow path, but it is not required by the supplied WanVideoWrapper workflow.

Repository layout

HighNoise/
  GingerLabs_Wan2.2_T2V_A14B_High_FP8_E4M3FN_SCALED.safetensors
  GingerLabs_Wan2.2_T2V_A14B_High_Q8_0.gguf
  GingerLabs_Wan2.2_T2V_A14B_High_Q6_K.gguf
  GingerLabs_Wan2.2_T2V_A14B_High_Q5_K_M.gguf
  GingerLabs_Wan2.2_T2V_A14B_High_Q4_K_M.gguf
LowNoise/
  GingerLabs_Wan2.2_T2V_A14B_Low_FP8_E4M3FN_SCALED.safetensors
  GingerLabs_Wan2.2_T2V_A14B_Low_Q8_0.gguf
  GingerLabs_Wan2.2_T2V_A14B_Low_Q6_K.gguf
  GingerLabs_Wan2.2_T2V_A14B_Low_Q5_K_M.gguf
  GingerLabs_Wan2.2_T2V_A14B_Low_Q4_K_M.gguf
workflows/
  GingerLabs_T2V_Workflow.json

Download example

hf download GingerLabsPlatform/GingerLabs-Wan2.2-T2V-A14B \
  HighNoise/GingerLabs_Wan2.2_T2V_A14B_High_Q5_K_M.gguf \
  LowNoise/GingerLabs_Wan2.2_T2V_A14B_Low_Q5_K_M.gguf \
  workflows/GingerLabs_T2V_Workflow.json \
  --local-dir GingerLabs-Wan2.2-T2V-A14B

Replace Q5_K_M with the same desired variant in both model filenames.

Resolution, duration, and memory tips

• Wan frame counts should follow 4n+1, such as 33, 49, 65, or 81 frames.
• At 16 fps, 33 frames is about 2.1 seconds; 49 is about 3.1 seconds; 81 is about 5.1 seconds.
• Use 480 × 832 portrait for a faster first test. The supplied 720 × 1280 preset is the higher-resolution portrait option.
• Increasing resolution, frames, or steps raises memory use and runtime.
• If VAE decode runs out of memory, enable tiling on WanVideoDecode.
• If text encoding runs out of VRAM, change the text-encode device from GPU to CPU; this is slower.
• VideoHelperSuite saves workflow metadata by default, which can include prompts and settings. Disable metadata saving when that information is sensitive.

Quantization details

The FP8 files store 400 selected attention and feed-forward matrices as E4M3FN with per-matrix FP32 scale weights. The other 695 source tensors are preserved as FP32.

The GGUF files were converted with ComfyUI-GGUF and its patched llama.cpp toolchain. The five-dimensional patch_embedding.weight tensor was restored after quantization for Wan/ComfyUI compatibility.

Validation

All ten weights passed structural validation and 12 CUDA load/unload checks:

• FP8 through native ComfyUI and ComfyUI-WanVideoWrapper
• GGUF through ComfyUI-GGUF
• 1,496 tensors in each FP8 file
• 1,095 unique tensors in each GGUF file
• Restored five-dimensional GGUF patch-embedding tensor
• No missing or duplicate model keys

The reference workflow also passed graph-integrity checks and imported without schema errors in the local ComfyUI frontend with WanVideoWrapper 1.3.6. Validation remains load-level only: no full video generation, visual-quality comparison, or performance benchmark was performed.

See manifest.json for repository-relative provenance and validation records, and checksums.sha256 for weight hashes.

Limitations

• These checkpoints are T2V-only. Do not use them as Wan 2.2 I2V, TI2V, FLF2V, S2V, Animate, or VACE models.
• Quantization can change output quality, prompt adherence, motion, and numerical behavior.
• Lower-bit variants generally reduce memory requirements at a greater potential quality cost.
• Results depend on the workflow, sampler, resolution, frame count, and supporting model assets.

Attribution and terms

This release is based on Wan-AI/Wan2.2-T2V-A14B, whose repository identifies the base model as Apache-2.0 licensed. Quantization and compatibility tooling are credited in NOTICE.md.

No new license claim is made by this model card. Use and redistribution remain subject to the terms applicable to the base model and GingerLabs modifications.