WanAnimatePlus Sampler

Easy Sampler exists so you don't have to look at this node. WanAnimatePlus Sampler is the full settings surface for the fork's sampling engine - and if you're running Bernini or SCAIL-2, you need some of what it exposes, because the Easy node doesn't reach the guidance modes those models depend on.

The core is familiar. model, image_embeds, steps (30), cfg (6), shift (5), seed, force_offload, scheduler (unipc default, 21 choices), plus text_embeds and samples (init latents for v2v) with denoise_strength and add_noise_to_samples for controlling how much of that init survives. context_options brings in windowing for long clips. batched_cfg batches cond and uncond for speed at the cost of memory - try it on a 24GB card, skip it on a 12GB one.

The part that makes it worth it: guidance modes. guidance_mode is where Bernini lives. Options: cfg (standard), apg (single-condition APG), apg_chain (image-reference APG chain), and cfg_chain (Bernini chained CFG). APG - adaptive parallel guidance - is the modern replacement for plain CFG, and the fork implements it with a full parameter set: apg_eta (parallel/orthogonal balance, 0.5 default), apg_momentum, apg_norm_threshold (L2 clipping, 50 default), and three omega values (apg_omega, apg_omega_I, apg_omega_TI) plus chain omegas (chain_omega_V, chain_omega_I, chain_omega_TI). The Bernini node literally outputs a recommended_guidance string telling you which of these to set per task - r2v wants apg_chain with apg_omega_I=4.5/apg_omega_TI=4.0, rv2v wants cfg_chain with the chain omegas. Defaults are sensible; follow the cheat-sheet.

Long-form and loop tools. riflex_freq_index (0 = disabled, default 6 when on) is RIFLEX - it lets new frames generate after the clip without looping back, which is the anti-repetition tool for extended takes. loop_args (from a loop-args node) handles loop-mode chunking. start_step/end_step slice the sampling schedule for two-stage tricks, and rope_function (comfy default) picks a compilable RoPE implementation - comfy_chunked trades some peak VRAM. The remaining inputs (feta_args, cache_args, slg_args, experimental_args, sigmas, unianimate_poses, fantasytalking_embeds, uni3c_embeds, multitalk_embeds, freeinit_args) are all optional bundles for specific models; ignore what you don't use. (flowedit_args is there but its tooltip is blunt: "FlowEdit support has been deprecated.")

Outputs. samples (latents for WanAnimatePlus Decode) and denoised_samples (pre-VAE denoised output).

Install via ComfyUI Manager (search "WanAnimatePlus") or git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus into custom_nodes, restart. Start on Easy Sampler if you're on plain Wan Animate; graduate here the day a Bernini or SCAIL-2 workflow needs apg_chain. The fork-wide chain rule applies - and when CFG 6 with a speed LoRA gives you melted output, remember the distilled-model rule: CFG toward 1.0.