Comfyui-anima-sampler

ComfyUI Anima Flow Corrective Sampler

Custom ComfyUI sampler nodes for Anima / Cosmos-style rectified-flow image models. It packages the official Anima Diffusers FlowMatch Euler schedule by default, with optional UniPC and PC3 solver variants for A/B testing in ComfyUI.

This is an independent implementation, not an official NVIDIA or CircleStone Labs release. The design is aligned with public Cosmos / Cosmos Predict2.5 pipeline ideas such as rectified-flow scheduling and UniPC-style predictor-corrector sampling.

The default profile follows the public Anima Diffusers scheduler config:

solver        = flow_euler
schedule      = flow_diffusers_linear_shift
flow_shift    = 3.0
steps         = 30
cfg           = 5.0
cfg_mode      = const
final_clean_pass = false

The goal is to improve prompt structure, spatial relationships, and detail stability while keeping the node surface small enough for daily use.

Update

May 31, 2026:

• The packaged default now uses 30 steps, CFG 5.0, constant CFG, flow_euler, flow_diffusers_linear_shift, flow_shift 3.0, and no final clean pass.
• Added flow_diffusers_linear_shift, mirroring the non-dynamic-shifting FlowMatchEulerDiscreteScheduler sigma grid used by Anima-Base-v1.0-Diffusers.
• Added two explicit Diffusers-grid experimental solvers: flow_unipc2_diffusers_x0 and flow_pc3_diffusers_damped. Both keep the existing solver math but use conservative low-sigma tail handling on the Diffusers timestep grid.
• Kept the previous enhanced profile selectable: flow_unipc2_x0 + flow_rf_linear_shift + flow_shift 5.0.
• Simplified the released node surface to Anima Flow Corrective Sampler and Anima Flow Settings.

Example Output

The files below were generated from the same prompt family and are included as single-image examples instead of a four-way grid. Import examples/workflows/anima_better_sampler.json in ComfyUI to inspect the example workflow.

| UniPC, linear shift5, const CFG 7 | PC3, linear shift5, const CFG 7 | | --- | --- | | <img src="https://raw.githubusercontent.com/KeithZ117/Comfyui-anima-sampler/main/examples/comparison/unipc_linear_shift_cfg7.jpg" alt="UniPC linear shift CFG 7 example" width="260"> | <img src="https://raw.githubusercontent.com/KeithZ117/Comfyui-anima-sampler/main/examples/comparison/pc3_linear_shift_cfg7.jpg" alt="PC3 linear shift CFG 7 example" width="260"> |

| er_sde + simple, CFG 4.5 | er_sde + simple, CFG 7 | | --- | --- | | <img src="https://raw.githubusercontent.com/KeithZ117/Comfyui-anima-sampler/main/examples/comparison/er_sde_simple_cfg45.jpg" alt="er_sde simple CFG 4.5 example" width="260"> | <img src="https://raw.githubusercontent.com/KeithZ117/Comfyui-anima-sampler/main/examples/comparison/er_sde_simple_cfg7.jpg" alt="er_sde simple CFG 7 example" width="260"> |

Nodes

• Anima Flow Corrective Sampler: the main sampler node.
• Anima Flow Settings: optional advanced controls for solver and CFG tuning.

The sampler works without connecting Anima Flow Settings; the tested defaults are built in. Connect the settings node only when you want to tune advanced parameters.

The sampler outputs both LATENT and IMAGE. Connect a VAE to the optional vae input when you want the image output decoded directly from the sampler. Leave vae disconnected when you only need the latent output.

ramp cfg starts guidance low and smoothly raises it to the selected cfg. With the default cfg=5, it starts near 4.5, keeps that low guidance through the early high-noise phase, and reaches 5 before the tail/detail phase.

Install

Install with ComfyUI-Manager when the node is available in the registry, or install directly from GitHub.

GitHub install:

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

Manual install:

Copy this repository into ComfyUI's custom_nodes directory:

ComfyUI/custom_nodes/Comfyui-anima-sampler

Then restart ComfyUI.

This repository does not include model weights. Install Anima and its required text encoder / VAE files according to the official Anima model card:

• https://huggingface.co/circlestone-labs/Anima

Recommended Use

Use Anima Flow Corrective Sampler in place of a normal sampler node.

Everyday controls:

• steps: default 30
• cfg: default 5.0
• cfg_mode: recommended const or ramp cfg
• flow_solver: default flow_euler
• flow_schedule: default flow_diffusers_linear_shift
• flow_shift: default 3.0 and used by shift-aware schedules such as flow_diffusers_linear_shift, flow_cosmos_rf_tail, and flow_rf_linear_shift
• denoise
• add_noise
• optional vae for direct image output

flow_cosmos is now the pure Cosmos RFlow-shaped schedule. Use flow_cosmos_rf_tail when you want the shifted RF-tail path controlled by flow_shift.

flow_cosmos_rho7 is the pure rho7/Karras-style Cosmos schedule. Connect Anima Flow Settings and enable flow_rho7_tail_auto to add the previous RF-tail-auto modification on top of rho7.

flow_diffusers_linear_shift mirrors the non-dynamic-shifting FlowMatchEulerDiscreteScheduler path used by the Anima Diffusers release. The packaged default uses flow_shift 3.0 and no final clean pass.

flow_rf_linear_shift follows the newer Cosmos Predict2.5 normalized RF linear inference grid with the same shift formula used by their FlowUniPC scheduler. It uses flow_shift directly; flow_shift 5.0 matches their published default shape. When this schedule is selected without connecting Anima Flow Settings, the sampler automatically disables final_clean_pass to match Cosmos 2.5's default "walk to terminal zero" behavior.

flow_rf_linear_s_tail_shift5 is a fixed shift-5 extension of flow_rf_linear_shift. It keeps the early section close to linear shift5, then uses an S-shaped sigmoid tail near the final 30-step region to enter low-noise refinement more smoothly. The external flow_shift value is ignored by this schedule because shift5 is baked into the preset name.

Current Default

The packaged default now follows the official Anima Diffusers-style path: flow_euler + flow_diffusers_linear_shift + flow_shift 3.0 + const cfg 5.0, with no final clean pass when the settings node is disconnected.

Official reference combinations:

• Cosmos2: AB2 x0/denoised solver + Karras/rho7 sigmas (sigma_max 80.0, sigma_min 0.002, rho 7) + constant CFG + final clean pass at the last non-zero sigma.
• Anima Diffusers: FlowMatchEulerDiscreteScheduler (shift 3, use_dynamic_shifting false, stochastic_sampling false) + Euler stepping, walking to terminal zero.
• Cosmos 2.5: FlowUniPC order 2 + normalized RF linear shift schedule (shift 5 in the released configs) + constant CFG, walking to terminal zero without the old Cosmos2-style final clean pass.

The previous enhanced profile remains selectable as flow_unipc2_x0 + flow_rf_linear_shift + flow_shift 5.0 + const cfg 7.0.

flow_pc3_damped + flow_cosmos and flow_cosmos_rf_tail + flow_shift 5.0 remain available as explicit alternatives from previous testing.

flow_cosmos_rf_tail preserves the old shifted RF-tail behavior. flow_cosmos_rho7 remains available as a quality/reference baseline. flow_rf_linear_shift is available for testing the newer Cosmos 2.5 default linear+shift schedule. flow_rf_linear_s_tail_shift5 is available as the fixed shift5 linear+S-tail variant. simple remains available for comparison with native ComfyUI-style schedules.

flow_unipc2_x0 is available as an Anima-adapted FlowUniPC solver. It applies the Cosmos 2.5 UniP/UniC BH predictor/corrector structure directly to ComfyUI's denoised/x0 output instead of treating it as Cosmos 2.5 velocity output. The settings node exposes the original-style UniPC controls for solver order, bh1/bh2, lower-order final, disabled early correctors, and dynamic thresholding.

flow_unipc2_diffusers_x0 reuses the same UniPC math with a conservative Diffusers-grid policy: order-2 body steps, first-order low-sigma tail steps, and tail corrector skips where the official Diffusers grid has very large lambda gaps.

flow_pc3_diffusers_damped reuses PC3 in the stable body of the Diffusers grid and skips endpoint correction in the low-sigma tail. This keeps the old flow_pc3_damped behavior unchanged while making the Diffusers-grid experiment explicit.

flow_ab2 is available as a one-model-call Adams-Bashforth 2 solver. It uses the previous x0 prediction after an Euler warmup step, matching the residual x0 AB2 idea in the Cosmos reference scheduler while using this sampler's normalized RF time.

The final clean pass is optional. When enabled, the integration loop stops at the last non-zero sigma instead of taking the appended terminal zero interval, then asks the model for one more x0/denoised prediction at that same non-zero sigma. This matches the explicit clean pass used in the Cosmos reference pipelines more closely than cleaning a terminal-zero state. The disconnected daily defaults for flow_diffusers_linear_shift, flow_rf_linear_shift, and flow_rf_linear_s_tail_shift5 leave it off so the linear RF presets walk to terminal zero.

Development

Run the local test suite from the repository root:

python -m unittest discover -s tests

References

• NVIDIA Cosmos Predict2: https://github.com/nvidia-cosmos/cosmos-predict2
• NVIDIA Cosmos Predict2.5: https://github.com/nvidia-cosmos/cosmos-predict2.5
• UniPC paper: https://arxiv.org/abs/2302.04867

License

This sampler code is released under the MIT License. See LICENSE.

Model weights are not included and are not relicensed by this project. The official Anima model card currently lists Anima under the CircleStone Labs Non-Commercial License and notes that it is also subject to the NVIDIA Open Model License Agreement for derivative Cosmos models.