Anima Checkpoint Loader (BSS)

Let's start with the honest version of the pitch: Anima is slow for a 2B model. The community consensus numbers are roughly 10 seconds per image on a 5090 and over a minute on a 4060, and it wants 30–50 steps on top of that. Speed was the model's one genuinely annoying problem, and this node is the whole-checkpoint answer to it.

Anima Checkpoint Loader (BSS) is the "I have one .safetensors file that contains the model, the CLIP, and the VAE" loader from the ANIMA_BOOSTER pack. It does exactly what the built-in CheckpointLoaderSimple does - it reads a full Anima checkpoint out of your checkpoints folder - and then it does two extra things you'd otherwise have to wire up yourself: it applies SageAttention and torch.compile to the model, and it patches a known Anima quirk so your sampler doesn't crash.

How it works

Under the hood it's comfy.sd.load_checkpoint_guess_config(), the same call ComfyUI itself uses. Then three patches happen on a cloned copy of the model (the original stays untouched in ComfyUI's cache):

• extra_conds patch. Anima's diffusion model has a preprocess_text_embeds step, and if the raw text embeddings skip it you hit a shape mismatch - the code comments call it the "1024 vs 2048" crash in KSampler. The loader forces the preprocessing so that crash never happens.
• SageAttention. A quantized attention kernel that's significantly faster than stock attention on Ampere (RTX 30xx) and newer GPUs. The auto mode lets SageAttention pick the best kernel. If the library isn't installed, it silently falls back to PyTorch's built-in SDPA - no error, no crash.
• torch.compile. JIT-compiles the transformer blocks one at a time via ComfyUI's set_torch_compile_wrapper. Individual-block compilation is friendlier to LoRAs than compiling the whole model, and it's the difference between the ~2.5× and ~3.5× tiers of the README's speed table.

One thing to know from the source: the loader does not force fp16, even though the README's speed table is labeled "fp16". Anima checkpoints are stored in bfloat16, and forcing fp16 produces black images due to a value-range mismatch. "Default" here means "leave the dtype alone," and that's the correct setting.

The inputs that matter

Only three, and two of them have sane defaults:

• ckpt_name - pick your Anima .safetensors from the checkpoints folder.
• sage_attention - auto (default) or disabled. Leave it on auto; on a GPU older than Ampere it disables itself and logs a warning rather than crashing.
• torch_compile - off by default. Turn it on for the big speedup, but only after reading the troubleshooting below.

The outputs are the usual trio: model, clip, and vae, which wire straight into your sampler, text encoder, and decoder like any checkpoint loader.

Installing it

The pack's requirements.txt is deliberately empty - every heavy dependency (SageAttention, Triton) is optional, and the nodes degrade gracefully without them. So installation is just the pack itself:

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

or click Manager → Install via Git URL and paste that URL, then restart ComfyUI. You'll find the node under BSS/AnimaBooster in the node menu. On Windows, if you want SageAttention or torch.compile to actually work, you'll need triton-windows and a precompiled sageattention wheel installed into your python_embeded environment - the README points at sdbds/SageAttention-for-windows for those.

Where people get burned

• The PassManager::run failed crash. This is the big one: if torch_compile is on and KSampler dies with a Triton error about _attn_fwd, it's a known Triton/LLVM/toolchain compatibility issue on Ubuntu 24.04. Easiest fix is setting torch_compile to False - you keep the SageAttention + TeaCache speed. Alternatives:

  export TRITON_JIT_DISABLE_OPT=1
  rm -rf ~/.triton/cache ~/.cache/triton /tmp/torchinductor_*

• "It's stuck!" No it isn't - the first 2–3 generations with compile enabled are slow because Triton is compiling kernels. That's the warmup, not a hang.
• Black images. If you were forcing fp16 somewhere, stop. Leave the dtype at default.

It's not the node you need if you keep your Anima model as separate files (a standalone diffusion model plus separate CLIP and VAE) - that's what Anima Booster Loader (BSS) is for. But if you grabbed a full Anima checkpoint off CivitAI, this is the one-click drop-in.