search

Kosinkadink / ComfyUI-AnimateDiff-Evolved

Improved AnimateDiff for ComfyUI and Advanced Sampling Support
Apache License 2.0
2.76k stars 208 forks source link

readme

AnimateDiff for ComfyUI

Improved AnimateDiff integration for ComfyUI, as well as advanced sampling options dubbed Evolved Sampling usable outside of AnimateDiff. Please read the AnimateDiff repo README and Wiki for more information about how it works at its core.

AnimateDiff workflows will often make use of these helpful node packs:

Installation

If using ComfyUI Manager:

  1. Look for AnimateDiff Evolved, and be sure the author is Kosinkadink. Install it. image

If installing manually:

  1. Clone this repo into custom_nodes folder.

Model Setup:

  1. Download motion modules. You will need at least 1. Different modules produce different results.
    • Original models mm_sd_v14, mm_sd_v15, mm_sd_v15_v2, v3_sd15_mm: HuggingFace | Google Drive | CivitAI
    • Stabilized finetunes of mm_sd_v14, mm-Stabilized_mid and mm-Stabilized_high, by manshoety: HuggingFace
    • Finetunes of mm_sd_v15_v2, mm-p_0.5.pth and mm-p_0.75.pth, by manshoety: HuggingFace
    • Higher resolution finetune,temporaldiff-v1-animatediff by CiaraRowles: HuggingFace
    • FP16/safetensor versions of vanilla motion models, hosted by continue-revolution (takes up less storage space, but uses up the same amount of VRAM as ComfyUI loads models in fp16 by default): HuffingFace
  2. Place models in one of these locations (you can rename models if you wish):
    • ComfyUI/custom_nodes/ComfyUI-AnimateDiff-Evolved/models
    • ComfyUI/models/animatediff_models
  3. Optionally, you can use Motion LoRAs to influence movement of v2-based motion models like mm_sd_v15_v2.
    • Google Drive | HuggingFace | CivitAI
    • Place Motion LoRAs in one of these locations (you can rename Motion LoRAs if you wish):
      • ComfyUI/custom_nodes/ComfyUI-AnimateDiff-Evolved/motion_lora
      • ComfyUI/models/animatediff_motion_lora
  4. Get creative! If it works for normal image generation, it (probably) will work for AnimateDiff generations. Latent upscales? Go for it. ControlNets, one or more stacked? You betcha. Masking the conditioning of ControlNets to only affect part of the animation? Sure. Try stuff and you will be surprised by what you can do. Samples with workflows are included below.

NOTE: you can also use custom locations for models/motion loras by making use of the ComfyUI extra_model_paths.yaml file. The id for motion model folder is animatediff_models and the id for motion lora folder is animatediff_motion_lora.

Features

Upcoming Features

Basic Usage And Nodes

There are two families of nodes that can be used to use AnimateDiff/Evolved Sampling - Gen1 and Gen2. Other than nodes marked specifically for Gen1/Gen2, all other nodes can be used for both Gen1 and Gen2.

Gen1 and Gen2 produce the exact same results (the backend code is identical), the only difference is in how the modes are used. Overall, Gen1 is the simplest way to use basic AnimateDiff features, while Gen2 separates model loading and application from the Evolved Sampling features. This means in practice, Gen2's Use Evolved Sampling node can be used without a model model, letting Context Options and Sample Settings be used without AnimateDiff.

In the following documentation, inputs/outputs will be color coded as follows:

Gen1/Gen2 Nodes

① Gen1 ① ② Gen2 ②
- All-in-One node
- If same model is loaded by multiple Gen1 nodes, duplicates RAM usage.		 - Separates model loading from application and Evolved Sampling
- Enables no motion model usage while preserving Evolved Sampling features
- Enables multiple motion model usage with Apply AnimateDiff Model (Adv.) Node
image image
image image

Inputs

Gen2-Only Inputs

Gen2 Adv.-Only Inputs

Gen1 (Legacy) Inputs

Outputs

Gen2-Only Outputs

Multival Nodes

For Multival inputs, these nodes allow the use of floats, list of floats, and/or masks to use as input. Scaled Mask node allows customization of dark/light areas of masks in terms of what the values correspond to.

Node Inputs
image 🟨mask_optional: Mask for float values - black means 0.0, white means 1.0 (multiplied by float_val).
🟦float_val: Float multiplier.
image 🟩mask: Mask for float values.
🟦min_float_val: Minimum value.
🟦max_float_val: Maximum value.
🟦scaling: When absolute, black means min_float_val, white means max_float_val. When relative, darkest area in masks (total) means min_float_val, lighest area in massk (total) means max_float_val.

AnimateDiff Keyframe

Allows scheduling (in terms of timesteps) for scale_multival and effect_multival.

The two settings to determine schedule are start_percent and guarantee_steps. When multiple keyframes have the same start_percent, they will be executed in the order they are connected, and run for guarantee_steps before moving on to the next node.

Node
image

Inputs

Context Options and View Options

These nodes provide techniques used to extend the lengths of animations to get around the sweetspot limitations of AnimateDiff models (typically 16 frames) and HotshotXL model (8 frames).

Context Options works by diffusing portions of the animation at a time, including main SD diffusion, ControlNets, IPAdapters, etc., effectively limiting VRAM usage to be equivalent to be context_length latents.

View Options, in contrast, work by portioning the latents seen by the motion model. This does NOT decrease VRAM usage, but in general is more stable and faster than Context Options, since the latents don't have to go through the whole SD unet.

Context Options and View Options can be combined to get the best of both worlds - longer context_length can be used to gain more stable output, at the cost of using more VRAM (since context_length determines how much SD sampling is done at the same time on the GPU). Provided you have the VRAM, you could also use Views Only Context Options to use only View Options (and automatically make context_length equivalent to full latents) to get a speed boost in return for the higher VRAM usage.

There are two types of Context/View Options: Standard and Looped. Standard options do not cause looping in the output. Looped options, as the name implies, causes looping in the output (from end to beginning). Prior to the code rework, the only context available was the looping kind.

I recommend using Standard Static at first when not wanting looped outputs.

In the below animations, green shows the Contexts, and red shows the Views. TL;DR green is the amount of latents that are loaded into VRAM (and sampled), while red is the amount of latents that get passed into the motion model at a time.

Context Options◆Standard Static

Behavior
anim__00005
(latent count: 64, context_length: 16, context_overlap: 4, total steps: 20)
Node Inputs
image 🟦context_length: Amount of latents to diffuse at once.
🟦context_overlap: Minimum common latents between adjacent windows.
🟦fuse_method: Method for averaging results of windows.
🟦use_on_equal_length: When True, allows context to be used when latent count matches context_length.
🟦start_percent: When multiple Context Options are chained, allows scheduling.
🟦guarantee_steps: When scheduling contexts, determines the minimum amount of sampling steps context should be used.
🟦context_length: Amount of latents to diffuse at once.
🟨prev_context: Allows chaining of contexts.
🟨view_options: When context_length > view_length (unless otherwise specified), allows view_options to be used within each context window.

Context Options◆Standard Uniform

Behavior
anim__00006
(latent count: 64, context_length: 16, context_overlap: 4, context_stride: 1, total steps: 20)
anim__00010
(latent count: 64, context_length: 16, context_overlap: 4, context_stride: 2, total steps: 20)
Node Inputs
image 🟦context_length: Amount of latents to diffuse at once.
🟦context_overlap: Minimum common latents between adjacent windows.
🟦context_stride: Maximum 2^(stride-1) distance between adjacent latents.
🟦fuse_method: Method for averaging results of windows.
🟦use_on_equal_length: When True, allows context to be used when latent count matches context_length.
🟦start_percent: When multiple Context Options are chained, allows scheduling.
🟦guarantee_steps: When scheduling contexts, determines the minimum amount of sampling steps context should be used.
🟦context_length: Amount of latents to diffuse at once.
🟨prev_context: Allows chaining of contexts.
🟨view_options: When context_length > view_length (unless otherwise specified), allows view_options to be used within each context window.

Context Options◆Looped Uniform

Behavior
anim__00008
(latent count: 64, context_length: 16, context_overlap: 4, context_stride: 1, closed_loop: False, total steps: 20)
anim__00009
(latent count: 64, context_length: 16, context_overlap: 4, context_stride: 1, closed_loop: True, total steps: 20)
Node Inputs
image 🟦context_length: Amount of latents to diffuse at once.
🟦context_overlap: Minimum common latents between adjacent windows.
🟦context_stride: Maximum 2^(stride-1) distance between adjacent latents.
🟦closed_loop: When True, adds additional windows to enhance looping.
🟦fuse_method: Method for averaging results of windows.
🟦use_on_equal_length: When True, allows context to be used when latent count matches context_length - allows loops to be made when latent count == context_length.
🟦start_percent: When multiple Context Options are chained, allows scheduling.
🟦guarantee_steps: When scheduling contexts, determines the minimum amount of sampling steps context should be used.
🟦context_length: Amount of latents to diffuse at once.
🟨prev_context: Allows chaining of contexts.
🟨view_options: When context_length > view_length (unless otherwise specified), allows view_options to be used within each context window.

Context Options◆Views Only [VRAM⇈]

Behavior
anim__00011
(latent count: 64, view_length: 16, view_overlap: 4, View Options◆Standard Static, total steps: 20)
Node Inputs
image 🟩view_opts_req: View_options to be used across all latents.
🟨prev_context: Allows chaining of contexts.

There are View Options equivalent of these schedules:

View Options◆Standard Static

Behavior
anim__00012
(latent count: 64, view_length: 16, view_overlap: 4, Context Options◆Standard Static, context_length: 32, context_overlap: 8, total steps: 20)
Node Inputs
image 🟦view_length: Amount of latents in context to pass into motion model at a time.
🟦view_overlap: Minimum common latents between adjacent windows.
🟦fuse_method: Method for averaging results of windows.

View Options◆Standard Uniform

Behavior
anim__00015
(latent count: 64, view_length: 16, view_overlap: 4, view_stride: 1, Context Options◆Standard Static, context_length: 32, context_overlap: 8, total steps: 20)
Node Inputs
image 🟦view_length: Amount of latents in context to pass into motion model at a time.
🟦view_overlap: Minimum common latents between adjacent windows.
🟦view_stride: Maximum 2^(stride-1) distance between adjacent latents.
🟦fuse_method: Method for averaging results of windows.

View Options◆Looped Uniform

Behavior
anim__00016
(latent count: 64, view_length: 16, view_overlap: 4, view_stride: 1, closed_loop: False, Context Options◆Standard Static, context_length: 32, context_overlap: 8, total steps: 20)
NOTE: this one is probably not going to come out looking well unless you are using this for a very specific reason.
Node Inputs
image 🟦view_length: Amount of latents in context to pass into motion model at a time.
🟦view_overlap: Minimum common latents between adjacent windows.
🟦view_stride: Maximum 2^(stride-1) distance between adjacent latents.
🟦closed_loop: When True, adds additional windows to enhance looping.
🟦use_on_equal_length: When True, allows context to be used when latent count matches context_length - allows loops to be made when latent count == context_length.
🟦fuse_method: Method for averaging results of windows.

Sample Settings

The Sample Settings node allows customization of the sampling process beyond what is exposed on most KSampler nodes. With its default values, it will NOT have any effect, and can safely be attached without changing any behavior.

TL;DR To use FreeNoise, select FreeNoise from the noise_type dropdown. FreeNoise does not decrease performance in any way. To use FreeInit, attach the FreeInit Iteration Options to the iteration_opts input. NOTE: FreeInit, despite it's name, works by resampling the latents iterations amount of times - this means if you use iteration=2, total sampling time will be exactly twice as slow since it will be performing the sampling twice.

Noise Layers with the inputs of the same name (or very close to same name) have same intended behavior as the ones for Sample Settings - refer to the inputs below.

Node
image

Inputs

Iteration Options

These options allow KSamplers to re-sample the same latents without needing to chain multiple KSamplers together, and also allows specialized iteration behavior to implement features such as FreeInit.

Default Iteration Options

Simply re-runs the KSampler, plugging in the output of the previous iteration into the next one. At the dafault iterations=1, it is no different than not having this node plugged in at all.

Node Inputs
image 🟦iterations: Total amount of times KSampler should run back-to-back.
🟦iter_batch_offset: batch_offset to apply on each subsequent iteration.
🟦iter_seed_offset: seed_offset to apply on each subsequent iteration.

FreeInit Iteration Options

Implements FreeInit, which is the idea that AnimateDiff was trained on latents of existing videos (images with temporal coherence between them) that were then noised rather than from random initial noise, and that when noising existing latents, low-frequency data still remains in the noised latents. It combines the low-frequency noise from existing videos (or, as is the default behavior, the previous iteration) with the high-frequency noise in randomly generated noise to run the subsequent iterations. Each iteration is a full sample - 2 iterations means it will take twice as long to run as compared to having 1 iteration/no iteration_opts connected.

When apply_to_1st_iter is False, the noising/low-freq/high-freq combination will not occur on the first iteration, with the assumption that there are no useful latents passed in to do the noise combining in the first place, thus requiring at least 2 iterations for FreeInit to take effect.

If you have an existing set of latents to use to get low-freq noise from, you may set apply_to_1st_iter to True, and then even if you set iterations=1, FreeInit will still take effect.

Node
image

Inputs

Noise Layers

These nodes allow initial noise to be added onto, weighted, or replaced. In near future, I will add the ability for masks to 'move' the noise relative to the masks' movement instead of just 'cutting and pasting' the noise.

The inputs that are shared with Sample Settings have the same exact effect - only new option is in seed_gen_override, which by default will use same seed_gen as Sample Settings (use existing). You can make a noise layer use a different seed_gen strategy at will, or use a different seed/set of seeds, etc.

The mask_optional parameter determines where on the initial noise the noise layer should be applied.

Node Behavior + Inputs
image [Add]; Adds noise directly on top.
🟦noise_weight: Multiplier for noise layer before being added on top.
image [Add Weighted]; Adds noise, but takes a weighted average between what is already there and itself.
🟦noise_weight: Weight of new noise in the weighted average with existing noise.
🟦balance_multipler: Scale for how much noise_weight should affect existing noise; 1.0 means normal weighted average, and below 1.0 will lessen the weighted reduction by that amount (i.e. if balance_multiplier is set to 0.5 and noise_weight is 0.25, existing noise will only be reduced by 0.125 instead of 0.25, but new noise will be added with the unmodified 0.25 weight).
image [Replace]; Directly replaces existing noise from layers underneath with itself.

Samples (download or drag images of the workflows into ComfyUI to instantly load the corresponding workflows!)

NOTE: I've scaled down the gifs to 0.75x size to make them take up less space on the README. The updated workflows have included Context Options and Sample Settings connected. The Context Options (and FreeNoise) do nothing unless context windows are triggered.

txt2vid

Result
readme_00461
Workflow
workflow-txt2vid

txt2vid - (prompt travel)

Result
readme_00463
Workflow
workflow-txt2vid-travel

txt2vid - 32 frame animation with 16 context_length

Result
readme_00475
Workflow
workflow-txt2vid-32frames

txt2vid - 32 frame animation with 16 context_length + ContextRef

Compared to without ContextRef, this tries to make the rest of the animation be more similar to the first context window.

Result
readme_00474
Workflow
workflow-txt2vid-32frames-contextref

Old Samples (TODO: update all of these + add new ones SOON)

txt2img - 32 frame animation with 16 context_length (uniform) - PanLeft and ZoomOut Motion LoRAs

t2i_context_mlora_wf

aaa_readme_00094_

aaa_readme00095.webm

txt2img w/ latent upscale (partial denoise on upscale)

t2i_lat_ups_wf

aaa_readme_up_00001_

aaa_readme_up00002.webm

txt2img w/ latent upscale (partial denoise on upscale) - PanLeft and ZoomOut Motion LoRAs

t2i_mlora_lat_ups_wf

aaa_readme_up_00023_

aaa_readme_up00024.webm

txt2img w/ latent upscale (partial denoise on upscale) - 48 frame animation with 16 context_length (uniform)

t2i_lat_ups_full_wf

aaa_readme_up00009.webm

txt2img w/ latent upscale (full denoise on upscale)

t2i_lat_ups_full_wf

aaa_readme_up_00010_

aaa_readme_up00012.webm

txt2img w/ latent upscale (full denoise on upscale) - 48 frame animation with 16 context_length (uniform)

t2i_context_lat_ups_wf

aaa_readme_up00014.webm

txt2img w/ ControlNet-stabilized latent-upscale (partial denoise on upscale, Scaled Soft ControlNet Weights)

t2i_lat_ups_softcontrol_wf

aaa_readme_up_00017_

aaa_readme_up00019.webm

txt2img w/ ControlNet-stabilized latent-upscale (partial denoise on upscale, Scaled Soft ControlNet Weights) 48 frame animation with 16 context_length (uniform)

t2i_context_lat_ups_softcontrol_wf

aaa_readme_up00003.webm

txt2img w/ Initial ControlNet input (using Normal LineArt preprocessor on first txt2img as an example)

t2i_initcn_wf

aaa_readme_cn_00002_

aaa_readme_cn00003.webm

txt2img w/ Initial ControlNet input (using Normal LineArt preprocessor on first txt2img 48 frame as an example) 48 frame animation with 16 context_length (uniform)

t2i_context_initcn_wf

aaa_readme_cn_00005_

aaa_readme_cn00006.webm

txt2img w/ Initial ControlNet input (using OpenPose images) + latent upscale w/ full denoise

t2i_openpose_upscale_wf

(open_pose images provided courtesy of toyxyz)

AA_openpose_cn_gif_00001_

aaa_readme_cn_00032_

aaa_readme_cn00033.webm

txt2img w/ Initial ControlNet input (using OpenPose images) + latent upscale w/ full denoise, 48 frame animation with 16 context_length (uniform)

t2i_context_openpose_upscale_wf

(open_pose images provided courtesy of toyxyz)

aaa_readme_preview_00002_

aaa_readme_cn00024.webm

img2img

TODO: fill this out with a few useful ways, some using control net tile. I'm sorry there is nothing here right now, I have a lot of code to write. I'll try to fill this section out + Advance ControlNet use piece by piece.

Known Issues

Some motion models have visible watermark on resulting images (especially when using mm_sd_v15)

Training data used by the authors of the AnimateDiff paper contained Shutterstock watermarks. Since mm_sd_v15 was finetuned on finer, less drastic movement, the motion module attempts to replicate the transparency of that watermark and does not get blurred away like mm_sd_v14. Using other motion modules, or combinations of them using Advanced KSamplers should alleviate watermark issues.