Anima Turbo Per-Step Expert Loader

Anima is slow for its size - it's the community's most repeated complaint about the model, and the escape hatch is distillation. If you're on the per-step-expert variant of that path (a turbo student trained with per_step_expert=true), AnimaTurboPerStepExpertLoader is the only loader that will actually run it. A stock LoRA loader can't, and here's the kicker: it won't even tell you. It'll silently produce no adapter, because every module is skipped as "missing router."

How it works

This kind of turbo student keeps one shared lora_down projection per Linear but carries K separate up-heads - and head k is assigned to serve denoise step k. There is no router, no gate, no per-input decision. Selection is by a denoise-step counter, which is exactly the piece of state a normal LoRA loader has no concept of.

The node installs a diffusion_model pre-hook that advances the head index once per model forward, plus per-Linear forward hooks that add the active head's delta. The on-disk file uses the trainer's fused qkv/kv key layout, so the loader splits those back into ComfyUI's q/k/v keys at load - the same contract Hydra uses, inlined so this node has no cross-package dependency.

The contract you must honor

Two settings make or break this node, both from the tooltip and the source docstring:

• cfg=1.0 - the turbo contract. With CFG 1, ComfyUI runs exactly one model forward per denoising step, so forward i maps cleanly onto head i % K. CFG > 1 runs two forwards per step and desyncs the head selection.
• infer_steps set to the trained head count K. Run 40 steps against a 12-head student and the mapping falls apart.

Miss either and you'll get garbage or nothing, and the node's assumption logging is your only clue. The README for this pack doesn't document this node yet - it ships in the same repo and installs identically, but you're reading ahead of the docs.

Inputs and output

• model - an Anima-family MODEL.
• adapter - the per-step-expert student (metadata stamped ss_turbo_per_step_expert=1), from the ComfyUI/models/loras dropdown.
• strength - default 1, range -2 to 2, scaling the active step-head delta.

Output is a single MODEL socket.

Install

Same as the rest of the pack: ComfyUI Manager (search "Anima Adapter Loader") or

cd ComfyUI/custom_nodes
git clone https://github.com/sorryhyun/ComfyUI-Anima_lora-Adapter

then restart. No extra dependencies.

Worth a word of caution that has nothing to do with this node: the community's read on Anima's distilled path in general is that it buys speed at the cost of the flexibility that made you switch to Anima in the first place. Per-step-expert turbo is a refinement of that idea, not a repeal of it - if you're on a 4060 and the wait is unbearable it's worth a try, but don't be surprised if you end up back at 30 steps for the images that matter.