Anima Safe PAG

한국어 README

Anima Safe PAG is a ComfyUI node for applying a softer form of Perturbed Attention Guidance (PAG) to Anima/Cosmos/Predict2-style DiT models.

It creates a weak prediction by perturbing selected self-attention blocks, then guides the final CFG result away from that weak prediction. Compared with hard PAG, the attention perturbation is blended by perturbation_strength, so the effect can be tuned more gently.

Cond, uncond, and PAG predictions are calculated in the same calc_cond_batch pass. The extra PAG row is also kept padded outside the active start_percent / end_percent range, which keeps the batch shape stable for workflows that use cache, replay, or sampler-side optimization nodes.

Comparison

Installation

ComfyUI Manager:

Manager -> Install Custom Nodes -> search "Anima Safe PAG" -> Install

Manual installation:

cd ComfyUI/custom_nodes
git clone https://github.com/iljung1106/comfyui-anima-safe-pag.git

Restart ComfyUI after installation.

Recommended Settings

scale: 4.0
block_indices: 18
perturbation_strength: 0.75
head_indices:
start_percent: 0.0
end_percent: 0.7
rescale: 0.20
rescale_mode: full

For 28-block Anima checkpoints, start with block indices greater than 14. Later blocks usually affect local structure, line confidence, and small details more cleanly than early blocks. 18 is a balanced starting point. Good nearby tests are 16, 18, 20, and 18-20.

Parameters

scale

Strength of the PAG correction. Higher values push the result farther away from the perturbed prediction. Reduce it if edges become harsh or details start to break.

block_indices

Transformer blocks where self-attention is perturbed. Use a single index, a comma-separated list, or a range: 18, 18,20, 18-20. For common 28-block Anima checkpoints, values above 14 are recommended.

perturbation_strength

Blend amount between normal self-attention and the value/identity path. 0.0 leaves attention unchanged. 1.0 is closest to hard PAG. The default 0.75 keeps the weak prediction useful without making the perturbation too aggressive.

head_indices

Optional attention-head filter. Leave it empty for normal use. On common 2048-channel / 16-head Anima checkpoints, heads are indexed from 0 to 15, but selecting heads manually can make results less predictable.

start_percent / end_percent

Active sampling range expressed as progress, not raw sigma. 0.0 is the start of sampling and 1.0 is the final step. The default 0.0 to 0.7 applies Safe PAG from the early steps through the mid/late part of sampling.

rescale

Guidance rescale amount. It reduces excessive contrast, saturation, or energy growth after the PAG correction. 0.0 disables rescaling. The default 0.20 is mild.

rescale_mode

full rescales the full CFG result after Safe PAG guidance. partial rescales against the positive prediction plus Safe PAG guidance. full is the safer default.

Compatibility

Anima Safe PAG expects an Anima/Cosmos/Predict2-style model with diffusion_model.blocks and block-level self_attn.compute_attention.

Existing model_function_wrapper behavior is preserved where possible, and no separate post-CFG model evaluation is used. The node works well in workflows with Anima Layer Replay Patcher, SPEED, or similar optimization nodes that rely on stable model execution.