Wan-Animate-2 14B — GGUF
GGUF quantizations of Wan-AI/Wan2.2-Animate-2-14B, converted for ComfyUI with ComfyUI-GGUF.
Wan-Animate-2 is a character animation model that takes a reference image and a driving video and transfers the performance directly — no intermediate skeleton or motion extractor in between. This repo exists so it runs on 8-12 GB cards instead of requiring the full 30 GB bf16.
All original licensing terms and usage restrictions carry over from the base model.
⚠️ These require a custom loader node
Load these with Rebels W3A8 Loader, not the stock GGUF loader.
Wan-Animate-2 doesn't announce itself through tensor names. ComfyUI selects it from a config blob in the safetensors __metadata__ that sets model_type: "animate2" — and GGUF has no equivalent field, so any GGUF of this model loads as a plain Wan 2.1 I2V.
The failure is quiet, which is what makes it worth spelling out: the model loads, sampling runs, a video comes out with your reference character in it — and the driving video is completely ignored. Random motion, no performance transfer. Nothing errors.
The loader node supplies that value explicitly:
- Install ComfyUI-GGUF and Rebels W3A8 Loader
- Use GGUF Unet Loader + model_type (Rebels) in place of the stock Unet Loader (GGUF)
- Set
model_typetoanimate2 - Everything else in the standard Wan-Animate-2 workflow stays as it is
Confirm it took — the console prints on load:
[Rebels GGUF] loaded <file> | model_type=animate2 | model class=WAN_Animate2
model class=WAN_Animate2 is the proof. Anything else and you're in the silent-failure case above.
Getting good motion transfer
The reference image must roughly match the driving video's opening pose. This is the single biggest quality factor and it isn't a quantization thing — it applies to the full bf16 model too. A front-facing portrait against a side-on driving video loses context and the model falls back to invented motion. Frame the reference the way the performance starts.
Files
Converted from the official bf16 weights. Every file was verified after quantization to confirm the precision-sensitive layers were preserved.
| Quant | Size | Notes |
|---|---|---|
| Q8_0 | ~17.4 GB | Near-lossless reference tier |
| Q6_K | ~13.5 GB | Excellent quality |
| Q5_K_M | ~11.7 GB | Very good |
| Q4_K_M | ~9.8 GB | Recommended — best quality per GB |
| Q4_K_S | ~9.2 GB | Slightly smaller |
| Q3_K_M | ~8.0 GB | Usable, visible detail loss |
| Q2_K | ~5.9 GB | Smallest; quality drops sharply |
Download a single file rather than the whole repo:
hf download realrebelai/Wan-Animate-2-14B-GGUF Wan-Animate-2-14B-Q4_K_M.gguf --local-dir ComfyUI/models/diffusion_models
Required companion files
The GGUF replaces only the diffusion model. Everything else comes from Comfy-Org's repackaged repo: Comfy-Org/Wan-Animate-2
| File | Folder | Purpose |
|---|---|---|
umt5_xxl_fp8_e4m3fn_scaled.safetensors | models/text_encoders | Text encoder (use umt5_xxl_fp16.safetensors if you have the VRAM) |
Wan2_1_VAE_bf16.safetensors | models/vae | VAE |
clip_vision_h.safetensors | models/clip_vision | Reference image encoder — required |
lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16.safetensors | models/loras | Optional step-distill LoRA for fewer steps |
hf download Comfy-Org/Wan-Animate-2 split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors --local-dir ComfyUI/models
hf download Comfy-Org/Wan-Animate-2 split_files/vae/Wan2_1_VAE_bf16.safetensors --local-dir ComfyUI/models
hf download Comfy-Org/Wan-Animate-2 split_files/clip_vision/clip_vision_h.safetensors --local-dir ComfyUI/models
Node setup
- Unet Loader (GGUF) — select the
.gguf - Load CLIP —
umt5_xxl_fp8_e4m3fn_scaled.safetensors, typewan - Load VAE —
Wan2_1_VAE_bf16.safetensors - Load CLIP Vision —
clip_vision_h.safetensors - Reference image + driving video into the Wan Animate nodes, then sample as usual
Requires a recent ComfyUI with native Wan-Animate-2 support, plus the ComfyUI-GGUF custom node.
Base vs distillation
Upstream ships two variants. This repo is quantized from the {base | distillation} weights.
- Base — standard step count and CFG. Higher ceiling, slower.
- Distillation — few-step schedule with CFG 1.0. Much faster; use the settings the base model documents rather than normal Wan settings, or output quality suffers regardless of quantization.
Conversion notes
Converted with a streaming converter that never loads the full model into RAM, then quantized with llama-quantize.
These layers are preserved at high precision rather than quantized:
patch_embedding, time_embedding, time_projection, text_embedding, img_emb, head, rope
That matters more than it sounds. Timestep and conditioning projections carry very little redundancy — quantizing them corrupts every sampling step, and standard quantizers do not preserve them automatically. They are restored after quantization and verified per file.
The 5-D patch_embedding conv weight is stored with its original shape recorded so the loader restores it correctly.
Credits
- Wan-AI / Alibaba — Wan2.2-Animate-2-14B
- Comfy-Org — repackaged encoder, VAE and CLIP vision files
- city96 — ComfyUI-GGUF
Quantized by RealRebelAI · GitHub · X