Load Anima XPred Model

If you grabbed an Anima x-pred checkpoint and wired it into the normal Anima loader plus a KSampler, you already know what happened: static soup, and no amount of prompt tuning fixed it. That wasn't you. x-pred checkpoints don't predict velocity the way ordinary Anima flow-matching checkpoints do - they predict clean latent x, and the sampler has to read the velocity back out as v = (z − x_pred) / sigma. Feed an x-pred checkpoint to a velocity-expecting sampler and the semantics are just wrong. Load Anima XPred Model is the pack's fix for exactly that.

What it does

It's the first node in the three-node chain from leafmoone/ComfyUI-RUM-Anima-XPred, and it's the one doing the heavy lifting. In one shot it loads three files: the Anima DiT x-pred checkpoint from models/diffusion_models/, the Qwen3 0.6B text encoder from models/text_encoders/, and the Qwen-Image VAE from models/vae/. Out the other end comes a single RUM_ANIMA_XPRED bundle - a type that only the pack's own Load Anima XPred LoRA and Sample Anima XPred nodes will accept. That's the guardrail working: you literally can't plug this into a normal velocity sampler, because a KSampler has no idea what a RUM_ANIMA_XPRED is.

Anima is Circlestone's 2B DiT on NVIDIA's Cosmos-Predict2, with Qwen3-0.6B for text and the Qwen-Image VAE, so this is the same stack a standard Anima workflow uses - the pack just loads it with x-pred semantics. It runs the checkpoint through the same Anima DiT loader the regular workflow uses, and it vendors a copy of sd-scripts' Anima inference code inside the repo, so you don't need a separate sd-scripts install or any pip packages beyond what ComfyUI already ships.

The inputs that actually matter

The loader is dropdown-driven - no absolute path text boxes to fight:

• checkpoint, text_encoder, vae - the three dropdowns. Pick xpred-adapter-checkpoint.safetensors, qwen_3_06b_base.safetensors, and qwen_image_vae.safetensors. Use the adapter checkpoint, not xpred-train-state.pt - the README is explicit about that one. No separate base Anima DiT file needed; the x-pred checkpoint is an Anima DiT checkpoint, loaded directly.
• prediction_type - leave on x. That's the entire point of the pack. The v option is for when you point this node at a plain velocity-prediction Anima checkpoint instead.
• precision - bf16 is the default and is right for most NVIDIA cards. Drop to fp16 if bf16 isn't supported, or fp32 if you're debugging precision weirdness and have the VRAM to spare.
• attn_mode - flash is default and fastest on modern NVIDIA GPUs. If you hit an attention-related error, torch or xformers are the fallbacks for older cards and AMD.
• text_encoder_cpu, fp8, fp8_scaled - the memory levers. Anima runs on 6GB cards per community reports, but the Qwen3 encoder is a real chunk of RAM. Ticking text_encoder_cpu (and optionally fp8) is how you squeeze onto smaller setups.

Install

The repo is young - a single commit as of mid-2026 - so expect rough edges, but install is boring:

cd ComfyUI/custom_nodes
git clone https://github.com/leafmoone/ComfyUI-RUM-Anima-XPred
# restart ComfyUI

ComfyUI Manager also has it; search "ComfyUI-RUM-Anima-XPred". Then drop the three model files into their models/ folders as listed above, plus any LoRAs under models/loras/.

Where people get burned

• Wiring the output into anything but this pack's nodes. It won't even connect, which is the guardrail working.
• The wrong checkpoint file. xpred-train-state.pt is a training artifact; the loader wants the adapter checkpoint.
• Flash attention on unsupported hardware. Swap attn_mode before you blame the model.
• Missing files. The loader raises a clear "not found" error naming the file, so double-check the dropdowns match what's actually sitting in models/.

From here the bare workflow is Load Anima XPred Model → Sample Anima XPred → Preview/Save Image, with the LoRA loader slotted in between if you have one.