Loads an INT8 Anima DiT with an fp32 input to its int8 kernels, which ComfyUI's own loader leaves at bf16 (comfy-kitchen#289). Stock LoRA and adapter loaders…

Anima loader nodes for ComfyUI. AnimaAdapterLoader auto-detects LoRA, HydraLoRA multi-head with live σ-conditional and FeRA-style FEI-conditional routing, ChimeraHydra dual-pool routing (per-Linear content router + network-level FreqRouter on FEI+σ), and LoReFT residual-stream edits. AnimaFeraLoader handles FeRA (Yin et al., arXiv:2511.17979) — both the author-faithful networks.methods.fera format and the plan2 stacked_experts_global_fei format with global router on the latent's spectral energy + independent stacked experts per Linear. AnimaSoftTokensLoader handles SoftREPA-parameterization soft tokens (Lee et al., arXiv:2503.08250) — per-layer × per-timestep-bucket learned tokens spliced into the crossattn embedding inside the first n_layers DiT blocks. Three single-purpose nodes; chain via the MODEL socket.

Anima Adapter Loader

unet_name

stream

Loads an INT8 Anima DiT with an fp32 input to its int8 kernels, which ComfyUI's own loader leaves at bf16 (comfy-kitchen#289). Stock LoRA and adapter loaders chain after it.

Name	Type	Default	Description
unet_name	COMBO		An INT8 Anima file from models/diffusion_models/, such as anima-base-v1.0_int8_gptq.safetensors (huggingface.co/sorryhyun/anima-base-v1.0-int8).
streamopt	COMBO	fp32	DiT residual stream precision. fp32: the stream, modulation, q / k norms, RoPE, embedder and head run in fp32 around the INT8 block GEMMs. bf16: only the kernels' input is fp32.

Anima INT8 Loader

Inputs (2)

Outputs (1)