Anima Flow Settings

First thing to know: the sampler works fine without this node. Anima Flow Settings is the optional second half of the Comfyui-anima-sampler pack - a bundle of advanced controls that only does anything when you wire its settings output into the sampler's optional flow_settings input. The tested defaults are baked into the sampler itself, so if you're getting good images, you can ignore this node forever. It exists for the day you want to reproduce an official Cosmos schedule, chase a specific artifact, or do the A/B testing the pack is built around.

What it is, and why it doesn't sample anything

It's a settings-bundle node, not a sampler. Nothing goes in on the optional side, and the outputs are settings (an ANIMA_FLOW_SETTINGS struct the sampler reads) and summary (a string that describes the resolved profile). Wire that summary into a text display and you get a one-glance sanity check of what the sampler is actually about to do - the cheapest verification tool in the whole pack.

The knob pile breaks into groups. UniPC solver controls (flow_unipc_order up to 6, flow_unipc_solver_type bh1/bh2, flow_unipc_lower_order_final, flow_unipc_disable_corrector_first, plus dynamic thresholding with a ratio and max-value cap). PC3 controls (flow_pc3_gamma, flow_pc3_tolerance). An flow_er_order for the ER solver. A big CFG-shaping cluster (cfg_early_scale, cfg_early_ramp_end, cfg_peak_boost, cfg_bump_start, cfg_bump_end, cfg_beta_alpha/cfg_beta_beta, late_cfg_scale, late_cfg_start, and the two legacy-progress toggles). And scheduler-level switches (final_clean_pass, cosmos_sigma_max/min, flow_rho7_tail_auto, rf_endpoint_noise_refresh_enabled with strength and cutoff).

The ones that actually matter

Most of these are dead weight for a beginner, but three are worth understanding:

• final_clean_pass - the Cosmos-2-style move: stop at the last non-zero sigma instead of walking to terminal zero, then ask the model one more time at that same sigma. The default is off, which matches how both Anima Diffusers and Cosmos 2.5 ship. Flip it on only if you're comparing against old Cosmos-2 reference runs.
• flow_rho7_tail_auto - does nothing unless you've set flow_schedule = flow_cosmos_rho7 on the sampler. It re-adds the older RF-tail modification on top of the pure rho7/Karras-style Cosmos schedule.
• The CFG-shaping knobs - these tune what the sampler's cfg_mode does. ramp cfg on the sampler is a single option; the early-scale/peak-boost/bump knobs here let you reshape that curve by hand. If you're happy with the built-in modes, leave them.

There's a trap baked into the layout: a knob for a solver you're not using does nothing. flow_pc3_gamma is dead air while flow_solver is flow_euler, and none of the UniPC knobs matter if you're not on flow_unipc2_x0. The pack's own README documents three official reference combinations you can dial in and compare on a fixed seed - Cosmos 2 (flow_ab2 + rho7 sigmas + constant CFG + final clean pass), Anima Diffusers (Euler + shift 3 + terminal zero), and Cosmos 2.5 (flow_unipc2_x0 + RF linear shift 5). That's the intended use of this node: reproduce a reference, don't improvise.

Install and requirements

Same as the sampler node - it's the same pack. ComfyUI Manager (search "Comfyui-anima-sampler") or:

cd ComfyUI/custom_nodes
git clone https://github.com/KeithZ117/Comfyui-anima-sampler.git

Restart after cloning. The repo ships no weights and no extra pip dependencies; you still need the Anima base model, Qwen3 text encoder, and Qwen-Image VAE from the official HuggingFace card (circlestone-labs/Anima).

The honest troubleshooting advice

If you're here because output looks off, don't start twiddling. Change one variable at a time on a fixed seed - that's the only way a converging flow sampler tells you anything, and it's the discipline this whole pack was built around. The two legacy-progress toggles exist only to reproduce old behavior; leave them false unless you've worked out that you need them. And if you just want better-looking Anima output without a science project, the community's answer was never more sampler knobs - it was a latent upscale pass at 1.25 with denoise 0.25–0.35. Use this node when you're curious about the mechanism, not when you're frustrated.