Anima Booster Loader (BSS)

If you use Anima in ComfyUI for any serious amount of time, the same thought hits you around the tenth image: it is so slow for a 2B model. Circlestone's own card recommends 30–50 steps, and the community timings - over a minute on a 4060, ~10s on a 5090 - are the most complained-about thing about the model. Anima Booster Loader (BSS) is the flagship node of the ANIMA_BOOSTER pack, and it's the piece that answers that complaint in a single node: load the model, get SageAttention and torch.compile applied automatically.

The key word in its name is "Loader." This node loads the model only - a standalone Anima diffusion model file from your diffusion_models folder - and hands you a single MODEL output. It's the direct equivalent of ComfyUI's UNETLoader, not CheckpointLoaderSimple. If you want the whole model+CLIP+VAE in one shot, use the pack's Anima Checkpoint Loader (BSS) instead; this one assumes you're loading CLIP (Anima uses a Qwen3 text encoder) and VAE (Qwen-Image) separately.

How it works

The load path is comfy.sd.load_diffusion_model(), identical to stock ComfyUI. Then three things happen on a cloned model:

1. The extra_conds patch. Anima's preprocess_text_embeds gets forced before conditioning reaches the sampler, which is what prevents the "1024 vs 2048" dimension mismatch crash in KSampler.
2. SageAttention in auto mode - a fast quantized attention kernel on Ampere-and-newer GPUs, with a silent fallback to PyTorch SDPA if the library isn't installed. The code even checks your GPU's compute capability and auto-disables on RTX 20xx and older rather than risk a Triton crash.
3. torch.compile, per transformer block, via ComfyUI's set_torch_compile_wrapper.

Here's the part that trips people up because the README is marketing-loose about it: the README's speed table calls the baseline "fp16," but the loader deliberately does not force fp16. Anima is saved in bfloat16, and forcing fp16 causes black images from a value-range mismatch. "Default" means no dtype override - exactly what UNETLoader does - and that's the correct, crash-free choice. Treat the README's "fp16" as shorthand for "the optimized baseline," not an instruction.

The inputs

Just three:

• model_name - your Anima .safetensors from the diffusion_models folder.
• sage_attention - auto (default) or disabled. Leave it on auto; it safely does nothing if SageAttention isn't installed.
• torch_compile - false by default. Flip it for the big speed bump, after you've read the troubleshooting below.

Output is a single model (MODEL), which wires into Anima TeaCache (BSS) and then KSampler. The README's recommended chain is AnimaBoosterLoader (sage auto, compile on) → AnimaTeaCache (threshold 0.15, adaptive on) → KSampler.

Installing it

Nothing heavy to install for the pack itself - its requirements.txt has every library commented out, so the nodes run on stock PyTorch out of the box and only use SageAttention/Triton when present.

cd ComfyUI/custom_nodes
git clone https://github.com/BlackSnowSkill/ANIMA_BOOSTER.git

Then restart ComfyUI; the node lives under BSS/AnimaBooster. On Windows portable builds, making SageAttention actually work means installing triton-windows and a precompiled sageattention wheel into python_embeded (the README points at sdbds/SageAttention-for-windows). On Linux it's pip install sageattention --no-build-isolation with the CUDA toolkit present.

Where people get burned

• Triton PassManager::run failed. The known one on Ubuntu 24.04 when torch_compile is on. Set it to False - SageAttention plus TeaCache still gets you the 2.5–3.5× tier - or try export TRITON_JIT_DISABLE_OPT=1 and clearing ~/.triton/cache.
• Warmup confusion. With compile on, the first 2–3 generations are much slower while Triton builds kernels. Subsequent ones get the advertised 20–40%. Not a bug.
• "Where's my CLIP?" This node only outputs a model. If your graph goes straight to a text encoder expecting CLIP, you missed the Checkpoint Loader variant.

One honest caveat from the wider scene: SageAttention has a history of causing visual artifacts on some DiT models - the Z-Image Base was a famous example where it produced black outputs on release day. The auto mode here is reasonably careful, but if you ever see weird artifacts after turning this on, the first thing to try is sage_attention: disabled. That's a thirty-second test that rules out the whole class of problem.