Kosinkadink-ComfyUI-Advanced-ControlNet/examples/anima_lllite_v2

Anima LLLite validation workflows

This folder contains simple Advanced-ControlNet workflows for the two Anima LLLite v2 checkpoints: the five conditioning inputs documented for the v2 any-test-like model and the v2 inpainting model. It also includes vanilla parity and effect-mask validation workflows.

Simple v2 workflows

Type	Workflow	Input image	Checkpoint
Any - Grayscale A	anima_lllite_any_grayscale_a.json	anima_lllite_any_grayscale_a_control.png	anima-lllite-any-test-like-v2.safetensors
Any - Grayscale B	anima_lllite_any_grayscale_b.json	anima_lllite_any_grayscale_b_control.png	anima-lllite-any-test-like-v2.safetensors
Any - Lineart	anima_lllite_any_lineart.json	anima_lllite_any_lineart_control.png	anima-lllite-any-test-like-v2.safetensors
Any - HED scribble	anima_lllite_any_hed_scribble.json	anima_lllite_any_hed_scribble_control.png	anima-lllite-any-test-like-v2.safetensors
Any - PiDiNet scribble	anima_lllite_any_pidinet_scribble.json	anima_lllite_any_pidinet_scribble_control.png	anima-lllite-any-test-like-v2.safetensors
Inpainting	anima_lllite_inpainting.json	anima_lllite_v2_control.png	anima-lllite-inpainting-v2.safetensors

Download the selected input image from the table and save it under the displayed filename in ComfyUI/input. Load its workflow and queue it unchanged. Results are saved under ComfyUI/output/acn_anima_examples. Binary inputs and screenshots are linked externally instead of being committed to this repository.

These examples use default node names and do not connect the custom 28-layer Anima weights node. The inpainting workflow uses Anima LLLite Extras and Default Weights only because the model's source mask must be carried through cn_extras. Strength, start/end scheduling, effect masks, timestep keyframes, latent keyframes, and stacking remain available on the standard Advanced-ControlNet nodes.

The model author trained the v2 any-test-like checkpoint on five conditioning types: HED scribble, PiDiNet scribble, Grayscale A, Grayscale B, and lineart, all with heavy augmentation. The examples above exercise each input type using that one v2 checkpoint instead of the lower-quality Preview3 depth, pose, lineart, and scribble checkpoints.

The HED and PiDiNet controls were prepared with the comfyui_controlnet_aux HED and Scribble PiDiNet preprocessors. The HED soft-edge output was thresholded at 80 to make the documented white-on-black scribble representation. The model card names two grayscale generation patterns but does not define their individual construction; Grayscale A is the author's sample and Grayscale B demonstrates the documented inversion, contrast, and blur augmentation.

Requirements

Official Hugging Face repositories:

• circlestone-labs/Anima - base model, text encoder, and VAE
• kohya-ss/Anima-LLLite - Anima LLLite control models

Use ComfyUI commit 0f42ba514631 or later and place these files in the listed model folders. These are direct downloads from the official model repositories:

Download	ComfyUI model folder
anima-base-v1.0.safetensors	models/diffusion_models
qwen_3_06b_base.safetensors	models/text_encoders
qwen_image_vae.safetensors	models/vae
anima-lllite-inpainting-v2.safetensors	models/model_patches
anima-lllite-any-test-like-v2.safetensors (3-channel any-type model)	models/model_patches

The included workflows use Load Anima LLLite Model, so their LLLite files belong in models/model_patches, matching vanilla ComfyUI. For compatibility with existing Advanced-ControlNet LLLite workflows, the files may instead be placed in models/controlnet and loaded with Load Advanced ControlNet Model. The standard loader automatically distinguishes Anima checkpoints from older SDXL LLLite checkpoints.

Run the inpainting comparison

This validation workflow runs the inpainting model through vanilla ComfyUI and Advanced-ControlNet with identical inputs and sampling settings. It saves both decoded results, both latent tensors, and an absolute pixel-difference image.

1. Download anima_lllite_v2_control.png to ComfyUI/input.
2. Load anima_lllite_v2_inpaint_comparison.json in ComfyUI.
3. Queue the workflow without changing its settings.
4. Inspect ComfyUI/output/acn_anima_pr.

The control PNG contains a transparent edit region. ComfyUI's Load Image node provides that alpha channel as the source inpainting mask.

The vanilla branch passes the mask directly to Apply Anima LLLite. The Advanced-ControlNet branch passes it through Anima LLLite Extras, the cn_extras input on Default Weights, and weights_override on Apply Advanced ControlNet.

If an inpainting checkpoint reaches sampling without that source mask, Advanced-ControlNet raises an error that describes these connections instead of silently substituting an empty mask.

With the included seed and settings, the expected results are:

• Identical latent tensors with maximum and mean absolute differences of 0.0.
• Identical decoded PNG pixels.
• A completely black absolute_difference image.

Run the any-type effect-mask comparison

This workflow verifies the official 3-channel any-type checkpoint against vanilla ComfyUI and applies an Advanced-ControlNet effect mask to only the left half of the image. Its nodes retain their default names; the colored regions identify the comparison branches.

1. Download anima_lllite_v2_any_control.png and anima_lllite_v2_left_half_mask.png to ComfyUI/input.
2. Load anima_lllite_v2_any_effect_mask.json in ComfyUI.
3. Queue the workflow without changing its settings.
4. Inspect ComfyUI/output/acn_anima_any_mask.

The expected results are:

• Advanced-ControlNet full control and vanilla full control have identical latent tensors and decoded pixels, with maximum absolute difference 0.0.
• A completely black effect mask produces the no-control latent exactly.
• A completely white effect mask produces the unmasked full-control latent exactly.
• At the model token resolution, the included half mask is exactly 1.0 on the left and 0.0 on the right. Direct LLLite injection is therefore zero for every right-side token.
• In the final image, the masked right side is closer to the no-control baseline than full control: PSNR improves from 16.39 to 17.69, and SSIM improves from 0.554 to 0.600.

The final right half is not pixel-identical to the no-control image. This model patches self-attention Q, so controlled left-side tokens can influence right-side tokens through global self-attention and later diffusion steps. The effect mask guarantees local control injection, not hard image-space isolation after attention.