Deforum Nodes
Official Deforum animation pipeline tools that provide a unique way to create frame-by-frame generative motion art.
Nodes (69)
Nineteen kinds of noise, one node, and when you'd ever want them
Rebuild actual audio from a magnitude and phase pair
The quick-and-dirty beat detector that needs no parameters
The librosa-powered upgrade when the simple detector wobbles
The type adapter that lets 0–1 amplitude data talk to the rest of the pack
Add scheduled film-grain and anti-blur noise to every frame
Amplitude curves become motion schedules
Turn an amplitude series into readable text so you can see what the music did
The two-boolean gate for chaining conditions in your animation
The node that decides what kind of animation you're making
Spray a suffix onto every prompt in your animation
Warp a frame along the motion of a real video
Render different prompts in different regions of the frame
53 ready-made math expressions for your audio schedules, no trial-and-error
Canvas size, seeds, and the sampler
The reset switch that forces your animation back to frame zero
Stash a frame in the pack's cache so another part of the graph can grab it
Keep a latent around for reuse without dragging wires across the graph
Stash text across animation passes without a wire
How Deforum gets a buttery animation from a handful of renders
Run diffusion less, interpolate the rest
Stop your animation from slowly shifting color as it drifts frame to frame
Pick your anchor and color space
Compare two integers and get a boolean your graph can act on
Smoothly morph your prompt into the next one mid-animation
The \"safe\" ControlNet apply that makes your animation stick together
Depth warping that moves your camera through the scene
The schedule-driven sampler settings behind every frame
Turn 30 sparse frames into 120 smooth ones with Google's FILM
Compare two floats to drive the logic chain
Peek inside the animation loop's frame data
The 3D camera move that makes Deforum feel like flight
Pick a frame back out of the animation's memory
How Deforum hands a latent to the next frame
A numbered slot for text your loop needs later
Let a reference video steer your animation frame by frame
The settings that fuse your video into the animation
One number that controls how hard the video pushes
A boolean that picks between two images — and why that's useful
The \"safe\" inpainting conditioning that keeps your animation on-model
The stateful engine that drives every frame of a Deforum animation
The KSampler that takes its orders from the Iterator
The front door to vid2vid — feed any clip into the loop
The fiddly knobs that shape your animation's motion
The one-node inverter for conditional animations
Combine conditions so one true is enough
The keyframe script that tells your story frame by frame
Stop reading logs in the terminal, watch them in the UI
73 pre-built expressions for people who don't do math
Let the seed write your animation curves
See your animation curve before you render it
The most boring node in the pack, and why it exists
An experimental tuning knob for tiling workflows
The weirdest way to remix a CLIP's vocabulary
Smooth out your animation's jumps with optical flow
Deforum's Integrated Pipeline
The camera rig of your animation, in schedule form
Deforum's Unformatted Prompt Node
The Latent Passthrough That Keeps Cadence Workflows Alive
The node that turns your frame stream into an actual video file
When Your Animation Needs the Attack, Not the Loudness
Drive Your Animation From the Melody, Not the Noise
Isolate the Kick Drum, Then Drive the Zoom From It
The Other Half of the Spectrum Round Trip
One Number That Tells You How 'Bright' the Track Is
Cut the Scene When the Track Slows Down
Stop your audio-reactive animation from twitching like a strobe
Deforum for ComfyUI
Deforum animation nodes for ComfyUI — 2D/3D camera motion, cadence, hybrid video, schedules and prompt interpolation, driven frame by frame from inside a normal ComfyUI graph.

Verified against ComfyUI 0.44.2 and deforum-studio 0.2.0.
Installation
ComfyUI Manager
Look for Deforum Nodes by XmYx. The Manager installs requirements.txt,
which pulls in the deforum backend and everything else the nodes need.
Manual
cd ComfyUI/custom_nodes
git clone https://github.com/XmYx/deforum-comfy-nodes.git
Then install the dependencies into the same Python that runs ComfyUI:
python -m pip install -r ComfyUI/custom_nodes/deforum-comfy-nodes/requirements.txt
"The same Python" differs per install, and this is the single most common reason the nodes fail to load. Use the full path — then it does not matter which folder you run the command from:
| Install | Interpreter |
| --- | --- |
| ComfyUI Desktop | <install>\ComfyUI\.venv\Scripts\python.exe |
| ComfyUI portable | <install>\python_embeded\python.exe |
| Manual / venv | the python of the venv you launch ComfyUI with |
For example, on ComfyUI Desktop:
"D:\Comfy-Desktop\ComfyUI-Installs\Comfyui\ComfyUI\.venv\Scripts\python.exe" -m pip install -r "D:\Comfy-Desktop\ComfyUI-Installs\Comfyui\ComfyUI\custom_nodes\deforum-comfy-nodes\requirements.txt"
This step needs Git. requirements.txt pulls the deforum backend straight
from GitHub (deforum @ git+https://...), and pip shells out to git to do it.
Without Git installed and on PATH, pip fails. Either install
Git for Windows, or install the backend from
a zip instead, which needs no Git:
<python> -m pip install https://github.com/XmYx/deforum-studio/archive/refs/heads/main.zip
Do not run
pip install deforum. An unrelated project of that name exists on PyPI; it installs cleanly and then fails at import. The backend this pack needs is deforum-studio.
The Save Video node also needs ffmpeg on PATH. Without it the frames render
and the final .mp4 write fails.
Python 3.11 – 3.14 are supported. Nothing here pins or reinstalls torch, numpy or moviepy, so it will not disturb an existing CUDA, ROCm, DirectML or Apple Silicon install.
Example workflows
Load these from examples/ with Workflow → Open. Each one ships with a
checkpoint name you will need to point at your own model.
| Workflow | Model | Notes |
| --- | --- | --- |
| deforum_base.json | SDXL | 3D camera move with depth warping. The best starting point. |
| deforum_cadence.json | SDXL | Adds cadence frames between diffused frames for smoother motion. |
| deforum_krea2_turbo.json | Krea-2 Turbo | 8-step turbo model, cfg 1. |
| deforum_flux3_klein.json | FLUX.3 Image Klein (base) | 50 steps, cfg 4.0, euler / simple, with ModelSamplingFlux shift = ln(16). |
| deforum_guided.json | SDXL | Scheduled guide images — pin an image to a keyframe and keep pulling towards it. |
| deforum_integrated.json | SDXL | Whole animation inside one Deforum Single Sample node. |


Older workflows live in examples/legacy/; see the README there for what still works.
Running one
- Load the workflow and set Load Checkpoint (or the UNET/CLIP/VAE loaders) to a model you have.
- Queue the prompt. The first run renders frame 0.
- The Iterator node ships with
enable_autoqueueon, so the run drives itself until it reachesmax_frames. You can also queue manually, or use ComfyUI's own auto-queue. - Save Video writes an
.mp4to your output folder oncemax_framesframes are cached.
To restart an animation from frame 0, tick reset_counter and reset_latent on
the Iterator for one run, then untick them. Leaving them ticked pins the run to
frame 0 forever.
Model support
The Iterator's latent_type defaults to auto. Wire your model into its
model input and it reads the latent shape straight from ComfyUI, so new
architectures work without waiting for this pack to be updated.
| Family | Latent | Works with auto |
| --- | --- | --- |
| SD 1.5 / SD 2.x / SDXL | 4 ch, /8 | yes |
| SD 3 / SD 3.5 | 16 ch, /8 | yes |
| FLUX.1 | 16 ch, /8 | yes |
| FLUX.2 / FLUX.3 | 128 ch, /16 | yes |
| Qwen-Image, Krea-2 | 16 ch, /8 | yes |
| Stable Cascade | 16 ch, /42 | yes |
Leave model unconnected only if you want to pick the shape by hand from the
dropdown.
Distilled models (Krea-2 Turbo, SDXL Turbo, FLUX schnell, the
flux3-image-klein-* distilled checkpoints, …) want cfg_scale_schedule set to
0: (1) in Diffusion Parameters. At a higher cfg they burn out.
FLUX.3 Klein base is the opposite: flux3-image-klein-base-* is the CFG
variant and wants real guidance. Follow Black Forest Labs' own settings —
cfg_scale_schedule 0: (4.0), steps_schedule 0: (50), euler / simple,
and a ModelSamplingFlux between the UNET and the samplers with both shifts set
to ln(alpha) = ln(16) = 2.77259. deforum_flux3_klein.json is wired that way.
Note that ComfyUI's own bundled FLUX.3 template uses cfg 1 for the base file;
that is the distilled model's recipe and gives mush on the base weights.
Checkpoints without a text encoder. SD 3.5, FLUX and Qwen-Image are usually
shipped as a bare diffusion model. Load Checkpoint then returns no CLIP and you
get 'NoneType' object has no attribute 'tokenize'. Load the text encoders
separately with CLIPLoader / DualCLIPLoader and feed that CLIP into
Blend Conditionings, as deforum_flux3_klein.json does.
Guide images
Two different things, depending on what you want.
Seed the animation from an image. Encode it and feed the Iterator's
init_latent input:
LoadImage ─► VAEEncode ─► (deforum) Iterator [init_latent]
That replaces the starting noise on frame 0 only. How closely frame 0 follows it
is strength_schedule[0] in Diffusion Parameters.
Keep guiding later frames. Use Guide Image Schedule, between Frame Warp and the rest of the loop:
Frame Warp ─► Guide Image Schedule ─► Color Match ─► Add Noise ─► VAEEncode
▲
LoadImage ─► ImageBatch
It takes a batch of guide images and two schedules.
image_schedule picks which guide to use, by index into the batch.
0:(0), 60:(1) moves from the first guide to the second over 60 frames, and
with crossfade_guides on, fractional values blend between neighbouring guides
so the handover is smooth rather than a jump cut.
strength_schedule says how hard each frame is pulled towards it. 0 passes
the frame through untouched, 1 replaces it outright.
Pulse it, don't hold it. This node sits inside the feedback loop, so a
strength held across many frames compounds and the animation converges onto the
guide — at a steady 0.12 the guide has effectively taken over within a dozen
frames. What you usually want is a pulse at each keyframe, leaving the animation
free in between:
0:(1.0), 2:(0.0), 58:(0.0), 60:(0.6), 62:(0.0)
That pins the first guide at frame 0, lets the animation run, then pulls hard
towards the second guide at frame 60 and releases again. A sustained pull is
still useful for locking composition, but keep it around 0.02–0.05.
How far the animation drifts back afterwards depends on strength_schedule in
Diffusion Parameters — that is the denoise. At the low denoise deforum
usually runs (0.2–0.5) the model has little room to redraw, so after a strong
pulse the animation carries on from the guide rather than returning to what it
was doing. That is what pinning a guide to a keyframe means; if you want the
guide to read as a suggestion instead, use a smaller pulse or raise the denoise.
Guides are resized to the frame automatically, and an optional mask input
restricts the guide to part of the frame. The node also outputs the resolved
guide for the current frame, which is worth hanging a Preview on while you dial
the schedules in.
deforum_guided.json is a working example; swap its two LoadImage nodes for
your own images.
For conditioning-level guidance instead (ControlNet, IPAdapter), use Schedule Value: it reads any deforum schedule as a FLOAT for the current frame, so you can drive a ControlNet strength from the deforum timeline.
Schedules
Most parameter fields take a deforum schedule: frame: (expression), comma or
newline separated, with t as the frame number.
0: (1.0), 30: (1.02), 60: (1.0)
0: (1.0025+0.002*sin(1.25*3.14*t/30))
Prompts are frame: text, one per line:
0: a lone astronaut walking across a vast red desert, cinematic
30: the desert dissolving into glowing crystal spires
Quotes around the prompt are optional and are stripped if present.
Prompt timing and blending
A prompt scheduled at frame N is rendered exactly as written from frame N. To
crossfade into the next prompt, set prompt_blend_frames on the Iterator to the
number of frames the fade should take; the fade ends on the next prompt's
keyframe, so earlier prompts keep their timing. prompt_blend_curve picks
linear or exponential. 0 frames means no blending at all.
Troubleshooting
IMPORT FAILED / nodes missing from the menu. The node pack now prints the
missing module and the exact pip command for your interpreter. The usual cause is
installing into a different Python than the one ComfyUI runs.
No module named 'deforum'.
python -m pip install git+https://github.com/XmYx/deforum-studio.git
No module named 'timm.layers'. timm is older than 1.0. Reinstall the
requirements, or python -m pip install "timm>=1.0".
No module named 'moviepy.editor'. moviepy 2.x dropped that module. The
video nodes handle both 1.x and 2.x now, so update this pack.
'NoneType' object is not subscriptable in KSampler. The deforum cache was
empty. Branches that depend on a previous frame now stop quietly instead of
crashing, so this should no longer reach you.
Frying / harsh edges after a few frames. Lower amount_schedule and
noise_schedule, raise strength_schedule, and set color_coherence to LAB.
Running a model below its native resolution (SDXL under 1024) does this too.
Colours drift over the clip. color_coherence on ColorMatch Parameters
only does something if a Color Match node is actually in the graph, between
Frame Warp and Add Noise. deforum_krea2_turbo.json and
deforum_flux3_klein.json show the placement.
The animation never advances past frame 0. reset_counter / reset_latent
are ticked on the Iterator. If you are using cadence, check that
diffusion_cadence is smaller than max_frames — stepping by more than the
whole animation resets the counter on every queue.
Python integer -1 out of bounds for uint64. A -1 in seed_schedule
(deforum's "pick one for me") reached the noise generator. -1 is now resolved
to a random seed whatever seed_behavior is set to, so update this pack.
Optional companion nodes
None of these are required, and this pack no longer installs anything on your behalf:
- rgthree-comfy
- ComfyUI-AudioScheduler
- ComfyUI_IPAdapter_plus
- ComfyUI-VideoHelperSuite
- ComfyUI-Advanced-ControlNet
- ComfyUI_FizzNodes
- ComfyUI-Frame-Interpolation
Contributing
- Fork the repository on GitHub.
- Create a branch for your change.
- Commit with clear messages.
- Open a pull request.
License
MIT. See LICENSE.
Community and support
