Sample Anima XPred

The whole reason the other two nodes in this pack exist is so Sample Anima XPred can do its thing: run the actual sampling loop with x-pred semantics, decode the result through the Qwen VAE, and hand you a normal ComfyUI IMAGE. Regular Anima/FM samplers expect a velocity-prediction checkpoint; this one knows your model predicts clean latent x, so at every step it reads the Euler update direction back out as v = (z − x_pred) / sigma and integrates with that.

That one formula is the pack. It's what lets an x-pred checkpoint run at the default 10 steps and CFG 1.0, where ordinary Anima wants 30–50 steps at CFG 4–5. The distilled path costs flexibility - there's a real community argument about whether Anima's turbo/distilled variants are worth the quality loss - but if you've got an x-pred checkpoint, this is the sampler that was built for it.

How the sampling actually runs

It builds a shifted sigma schedule, seeds standard-normal noise latents (one per batch item, seed + i), encodes your prompt and negative through Qwen3, then steps through with heun or euler. Heun mode follows JLT's predictor-corrector structure - it evaluates the velocity, does a correction pass, then averages the two - and drops to a single Euler step for the final hop into σ=0. The eps_floor knob clamps σ during that (z − x_pred) / sigma division so you never divide by zero in the last step.

The settings you'll actually touch

• prompt / negative_prompt - Anima reads both Danbooru tags and natural language, and the default prompt is hatsune miku, 1girl, masterpiece, best quality. A fresh node will happily draw Miku at 1024×1024 before you type a word.
• steps - 10 by default, and that's the point of this path. Push to 15–20 if quality needs it; the field allows up to 200, but past ~30 you've missed the point.
• sampler - heun (default) buys better quality per step; euler is faster. Start on heun.
• cfg - 1.0 by default, meaning guidance is off. This is a CFG-1-style path and that's the right default. If you raise it, the negative_prompt suddenly matters.
• flow_shift - 3.0 is Anima's own default, the same shift knob other flow-matching samplers expose. Leave it alone unless you're chasing a composition-vs-detail tradeoff.
• width / height - 1024 default; both must be divisible by 8 or the node raises an error.

Less touched: seed (default 20260701, full 64-bit range), batch_size (1–16), eps_floor (1e-4 is fine), and offload_after_sample - keep that on. It shoves the model back to CPU and clears VRAM after each generation, which is what makes iterating on a 6GB card survivable.

Outputs

Two of them. image is a standard IMAGE tensor that goes straight into Preview Image or Save Image. latent is a raw RUM_ANIMA_LATENT dict carrying the final samples and sigmas - useful for debugging, or for feeding another RUM-node later.

Troubleshooting

• It came out as garbage / static. You almost certainly fed an x-pred checkpoint into a regular Anima sampler upstream. The pack's loader and this sampler are a matched pair - don't mix.
• "width and height must be divisible by 8." Bump the resolution to a multiple of 8.
• Non-finite latent errors - usually a precision edge. Try fp16/fp32 on the loader's precision, or nudge eps_floor up.
• Flash attention errors on the loader - switch attn_mode to torch or xformers before you blame sampling.

Install

Same as its siblings, all one repo:

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

or ComfyUI Manager → search "ComfyUI-RUM-Anima-XPred" → restart. Model files go in models/diffusion_models/, models/text_encoders/, and models/vae/ as the loader's dropdowns expect; this node itself needs nothing but the bundle from the loader.

The full workflow is three nodes - Load Anima XPred Model → Sample Anima XPred → Preview/Save Image, with the LoRA loader in the middle if you're using one. The pack is brand new and essentially unreported in the wild, so treat the defaults as starting points, not dogma: heun, 10 steps, CFG 1, 1024². Then explore.