Neo Turbo 加速 (Forge Neo)

Distilled speed LoRAs are the best deal in this hobby and the worst-documented. You download a Turbo or Lightning LoRA, you stack it on your normal workflow, and you get oversaturated, blown-out garbage - because the teacher's classifier-free guidance is already baked into the student. Running CFG 7 on top of it is applying the correction twice.

Anima's official Turbo LoRA is one of those. Neo ships a dedicated "Turbo" section that loads it and switches the whole sampler over in one move, and this node is that mechanism, moved into ComfyUI. The settings it enforces are not suggestions.

What the node actually does

Two things, and the second is the interesting one.

It loads your chosen LoRA model-only - no text encoder patch at all, which is the correct choice for a speed adapter. Then it returns a second output, turbo, a small package carrying a sampler name, scheduler, step counts and CFG.

Wire that turbo output into the Neo KSampler's turbo input and the sampler stops using its own widgets for those fields and uses the Turbo values instead. Step count is chosen by what the sampler is doing: steps for text-to-image, hires_steps for hires and img2img passes.

And because it's a wire and not a global switch, bypassing the node with the group's bypass control sends None down the wire, the LoRA passes through unapplied on the MODEL side, and the sampler falls back to whatever you set on its own widgets. That's the design: one node to toggle between "fast" and "slow" without rewiring anything.

The inputs

• model - your MODEL wire.
• lora_name - a dropdown populated from models/loras. The tooltip names the intended file: the official Anima Turbo LoRA from CircleStone Labs (Civitai model 2560840). It's a dropdown, so the file has to be in place before the node shows it.
• strength - default 1.0. The tooltip quotes the official note that a touch below 1 trades a little fidelity for variation, which is the standard complaint about distilled models: they produce suspiciously similar images across seeds. Dropping to 0.8–0.9 is the usual fix.
• sampler_name - default Euler, Neo's list.
• scheduler - default Simple. Compare that to the Neo KSampler's default of Automatic and take the hint: distilled models want the conservative schedules, not Karras.
• steps - default 10, tooltip says the official recommendation is 8–12. Do not "help" it by raising this to 20. A 4-step model run at 20 steps doesn't get better, it gets artifacts, because the trajectory was trained to be covered in a handful of big jumps.
• hires_steps - default 8, used for the second pass.
• cfg - default 1.0, and the tooltip says the quiet part out loud: at CFG 1 the negative prompt does nothing. That's not a bug you should work around by nudging CFG to 2; that's the distillation.

Outputs: MODEL into the Neo KSampler's model, and turbo into its turbo input. Both wires, or it does nothing.

Wiring

MODEL ─→ NeoTurbo ──MODEL──→ NeoKSampler.model
              └─────turbo──→ NeoKSampler.turbo

The sampler logs the Turbo settings it took over when it runs, which is the fastest way to confirm the wire is live: if you don't see that line, the turbo input is empty.

Install

ComfyUI Manager → ComfyUI Neo Sampler, or:

cd ComfyUI/custom_nodes
git clone https://github.com/wangjue520/ComfyUI-Neo-Sampler

Restart. The pack's three dependencies (lark, torchsde, scipy) auto-install into the Python that runs ComfyUI; if that fails the console prints the pip command. The LoRA itself is a separate download you have to put in models/loras yourself.

Reality checks

An empty lora_name dropdown is almost always a missing file, not a broken install - this node's list comes straight from the folder, and if the LoRA isn't there the widget has nothing to offer.

The bigger expectation-management point: a Turbo LoRA is a compromise, not a free lunch. Fewer steps means less time to refine detail, and every distillation method ever shipped loses something to get there. If your prompts relied on the negative prompt doing work, CFG 1 takes that away, and stacking three character LoRAs on top of a Turbo LoRA is where the fragility really shows - the same pattern people ran into with Z-Image Turbo's ecosystem. Use Turbo for drafts, wide shots and idea-mining, then re-run the keeper at normal steps with the LoRA bypassed. That's the workflow this node was built to make a two-click switch.