Deploy Wan 2.1

Cause: A distill/acceleration LoRA and a pre-distilled checkpoint are alternative acceleration paths — the rule is to disable the distill LoRA whenever the checkpoint is already pre-distilled. Stacking two conflicting few-step schedules degrades output.

Fix: Apply AT MOST ONE acceleration path. Note: base Wan DOES need the lightx2v LoRA for few-step (the correct single path); the double-distill hazard only arises with a pre-merged/FusionX-style Wan checkpoint that already bakes lightx2v/CausVid in.

Cause: fp8_e4m3fn_scaled checkpoints carry per-block scale factors that preserve dynamic range, matching bf16 quality for many modern diffusion/video models at ~half VRAM. Plain (unscaled) fp8_e4m3fn CAN visibly degrade — that's why the scaled repacks exist.

Fix: Prefer the 'scaled' fp8 variant; don't reflexively pick bf16. Do NOT generalize the claim to unscaled fp8.

Cause: Newer models replaced CLIP/T5 with a full LLM encoder. LTX-2 19B's LTXVGemmaCLIPModelLoader wants a DIRECTORY encoder: weights come from the non-gated Comfy-Org/ltx-2 repo, but config.json + tokenizer.model come from the GATED google/gemma-3-12b-it repo — without HF token + accepted license the sidecar 403s and the loader sees an empty list.

Fix: Supply HF token + accept the Gemma license so the gated config/tokenizer sidecar downloads into the encoder directory. Note LTX-2.3's default graph instead uses a single-file FP4 encoder (gemma_3_12B_it_fp4_mixed, non-gated, no tokenizer sidecar).

Cause: Loaders are picky three orthogonal ways: (1) PRECISION — Kijai LoadWanVideoT5TextEncoder wants its own bf16 umt5 build and rejects native ComfyUI's scaled fp8; native CLIPLoader(type=wan) accepts fp8. (2) PACKAGING — LTX LTXVGemmaCLIPModelLoader wants a directory encoder; LTXAVTextEncoderLoader wants a single-file fp4-mixed .safetensors. (3) EXACT NAME — t5xxl_fp8_e4m3fn.safetensors (4.89GB) and t5xxl_fp8_e4m3fn_scaled.safetensors (5.16GB) are distinct, not interchangeable.

Fix: Match the encoder file to the node's precision + packaging + exact filename; fuzzy substitution fails hard.

Cause: ComfyUI-GGUF's UnetLoaderGGUF reads a dedicated 'unet_gguf' folder key (fed from both unet and diffusion_models). models/unet is the reliably-covered location; the equivalent .safetensors goes in models/diffusion_models.

Fix: Route .gguf diffusion weights to models/unet (canonical, always covered). A GGUF in diffusion_models may not appear depending on ComfyUI/GGUF version — don't rely on it.

Cause: Wan 2.2 14B uses a two-model expert split — a high_noise AND a low_noise checkpoint (both ~fp8_scaled ~14GB) sampled in sequence. One 'model' is actually two files. (5B variant is a single ti2v model; i2v also needs clip_vision_h.)

Fix: Download BOTH the high_noise and low_noise 14B checkpoints (~28GB disk). VRAM stays ~one expert — they sample sequentially and ComfyUI swaps/unloads between them. Use umt5_xxl (not standard t5xxl or a CLIP); VAE differs by version (5B: wan2.2_vae, 14B: wan_2.1_vae).