Simple Load Anima

Anima is the anime model that actually made people switch - Circlestone Labs' 2B DiT on NVIDIA's Cosmos-Predict2 backbone, with a Qwen3-0.6B text encoder and a Qwen-Image VAE. And that's exactly why it's annoying to load with stock tools: it isn't a single checkpoint file. You need the diffusion model, the Qwen text encoder, the Qwen VAE, and then you have to decide precision and device placement for each. SimpleSyrup.SimpleLoadAnima collapses that whole setup into one node that outputs the standard model, clip, vae trio.

The author's framing is that loading a checkpoint used to mean choosing one file, and newer model families changed that. Anima is the clearest example in the pack: three separate components that all need to be the right version, or the model either fails to load or quietly prompt-encodes wrong.

Inputs, and what to do with them:

• diffusion_model - the Anima diffusion model file. This is the one you must provide yourself; the node won't download it.
• diffusion_weight_dtype - weight precision, default default. Lower precision cuts memory at the cost of marginally different results. Start on default and only drop it if you're VRAM-constrained.
• text_encoder - defaults to auto. Auto finds the expected Qwen text encoder on disk, and if it's missing, downloads the known, checksum-pinned support file with visible Comfy progress. You can also pick an explicit file if you have one.
• text_encoder_device - defaults to default (normal ComfyUI placement). Move it to CPU to save VRAM; prompt encoding just gets slower.
• vae - defaults to auto, resolving to the expected Qwen image VAE the same way.

The "auto" choices are the reason this node exists. Anima's Qwen components are checksum-pinned and revision-locked in the pack's catalog, so when auto resolves them it knows exactly which file it wants - it's not guessing at "some Qwen model." It finds it if present, downloads it if not, and validates against the pinned checksum. That's a much better failure mode than "the model loaded but colors are all wrong because some random Qwen VAE got picked."

One caveat worth carrying in from how Anima behaves generally: it's a 2B model that wants 30–50 steps, so per-step cost is low and wall-clock speed is mediocre on small cards (fine on a 5090, over a minute per image on a 4060). The loader doesn't change any of that - it just gets the right parts into the graph.

Install: the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup
cd SimpleSyrup && pip install -r requirements.txt

or ComfyUI Manager → search SimpleSyrup → Install → restart, then keep ComfyUI current - SimpleSyrup uses the v3 extension API. The auto-download path uses HuggingFace's download helpers, which the pack's requirements include (huggingface-hub). If downloads fail, check your network and disk space first; the checksum pin means a partial or corrupt download is detected rather than silently loaded.