Reforum: Deforum animations for ComfyUI

Reforum: Deforum animations for ComfyUI

Reforum (formerly "Deforum revived") brings Deforum animations back to ComfyUI: an unofficial, revived port of the Deforum animation engine — the one of the AUTOMATIC1111 extension — to ComfyUI 0.40+ (Python 3.13, torch 2.10, numpy 2). The classic Deforum look: frame-by-frame img2img feedback with keyframed 2D/3D camera motion, colour coherence, Perlin noise and anti-blur.

It is not affiliated with, nor endorsed by, the Deforum team. All credit for the engine goes to deforum and contributors; see Credits.

Why

The existing Deforum nodes for ComfyUI no longer load on recent ComfyUI (they rely on the old auto-queue, on a hook into ComfyUI's executor and on Deforum Studio, which pins numpy 1.x and an old diffusers). This package ports the engine itself, so it runs inside one node with ComfyUI's own samplers.

Nodes

| Node | What it does | |---|---| | Deforum Preset (revived) | The quick way: choose a recipe, write the scenes. Turbo veloce 3D and Turbo veloce 2D (SDXL Turbo + TAESDXL, 1 real step a frame, prompts that morph into each other, colours that follow every scene, a camera that changes direction by itself; about 0.7-0.9 s a frame at 768×432 on an RTX 3060) or Deforum classico (A1111) (the historical defaults). max_frames, W, H at 0 = the preset's. | | Deforum Settings (revived) | Every parameter, with the names of A1111 Deforum; the widgets start from the Turbo veloce 3D recipe (the historical A1111 values are the Deforum classico preset, and the defaults of a loaded A1111 settings file). Prompts as Deforum JSON. Any other parameter in extra_json. | | Deforum Load Settings File (revived) | Reads a settings file saved by A1111 Deforum (put it in ComfyUI's input folder) or the same JSON pasted, with optional overrides. Unsupported options found in the file are reported. | | Deforum Render (revived) | The engine. MODEL, CLIP, VAE, settings; optional init_image, video_init (Video Input mode), color_image (colour coherence "Image"), da3_model / moge_model / depth_maps (3D). The loop runs inside the node: one Queue renders the whole animation, with a preview of every frame. As Deforum did, every frame is also written as it is made to a folder in ComfyUI's output (Deforum_<timestring>/<timestring>_000000000.png …) together with <timestring>_settings.json (the same JSON A1111 Deforum wrote in its _settings.txt), which Load Settings File reads back (as well as A1111 .txt files) (the real seed is in it, also when you asked for -1). Outputs frames, depth, fps, folder. |

Then use ComfyUI's native nodes: CreateVideo + SaveVideo, and FrameInterpolate (RIFE/FILM) if you want. Examples (API format) are in examples/: turbo_veloce_3d/2d, klein4b_veloce_3d/2d, zimage_turbo_veloce_3d/2d (*_api.json) and the classic SD 1.5 ones. The fast examples have a Load Image on init_image (the first frame starts from that picture, kept as much as strength_schedule says): pick your image there, or bypass the node (Ctrl+B) to start from the prompt alone.

What is ported

• 2D motion: angle, zoom, translation x/y, transform centre, border replicate/wrap, perspective flip.
• 3D motion: translation x/y/z, rotation 3D x/y/z, fov, near/far, aspect ratio; the depth warp of Deforum (adapted from Disco Diffusion, PyTorch3D cameras). Depth from ComfyUI's native Depth Anything 3 or MoGe, re-estimated on every frame like Deforum did, or from your own depth maps.
• Schedules with Deforum's syntax and maths (t, max_f, s, sin, cos, where …), evaluated safely (no eval, no numexpr).
• Generation: strength, noise, contrast, cfg, steps scheduling, sampler scheduling, clip skip scheduling, noise multiplier, subseed; txt2img for the first frame; init image; cadence; Video Input mode.
• Colour coherence LAB / HSV / RGB / Image, grayscale; seed behaviour iter / fixed / random / ladder / alternate / schedule; Perlin and uniform noise; anti-blur (unsharp mask).
• Prompts: keyframed prompts, --neg, maths in backticks, global positive/negative prompts, A1111 AND.
• Extension: prompt_blend_frames morphs from one keyframed prompt to the next over N frames (the two prompts are mixed with moving weights); 0, the default, is Deforum's hard cut. color_coherence_anchor: scene releases the colour anchor around every prompt change and re-takes it from the new scene (Deforum keeps the first frame's colours for the whole animation, first_frame, the default). transition_strength lowers the strength smoothly around every prompt change (0, the default for A1111 files, = off). fps is on the Settings node.

Same behaviour: the order of the operations, the schedules, denoise = 1 - strength, and the A1111 step count for img2img (int(steps × denoise) + 1 steps on the tail of the schedule, i.e. KSamplerAdvanced, not KSampler's denoise). The functions were checked one by one against the original code (commit a7b829e): schedules, 3D warp and seeds are identical; 2D warp, colour coherence, noise and sharpening differ by about 0.2 levels (out of 255) or less on average (OpenCV → torch/numpy). The pictures are not identical to an A1111 render because ComfyUI's sampler noise is not A1111's.

Kept on purpose, because they are part of Deforum's look: the colour matching happens on BGR data passed to the RGB→LAB/HSV conversions; with perspective flip on, the 2D translation is ignored (a np.matmul quirk of the original). Changed on purpose: with cadence the last generated frame is in the output (Deforum dropped it); in Video Input mode the length is the length of the video, as in Deforum.

Speed

Warp, colour coherence, noise and anti-blur run on the GPU (torch), the PNGs are written in a background thread. For SDXL Turbo use the sampler Euler a SD Turbo (ComfyUI's SDTurboScheduler: steps are real steps from the noise level of the denoise, so 1 step works in img2img) with cfg 0:(1); a TAESDXL VAE (Load VAE → taesdxl) makes the VAE round trip of every frame almost free. diffusion_cadence is Deforum's own speed-up.

Long animations: the Render node also writes <timestring>.mp4 in the folder as the frames are made (save_video, on by default: there is a video also if you stop the render), and with keep_frames_in_memory off the frames go only to the folder and the node outputs only the last one, so thousands of frames do not fill the RAM.

Not ported (yet)

Hybrid video (composite and motion), masks and noise masks, optical-flow cadence and redo, diffusion redo, guided images (looper), checkpoint scheduling, resume, Interpolation mode, Parseq, ControlNet. The old MiDaS / AdaBins / LeReS / Zoe models are replaced by ComfyUI's native depth models.

Differences you may notice

• A1111's [word] de-emphasis is not understood by ComfyUI's text encoder; (word:1.2) works.
• A1111 AND is done with ComfyUI's conditioning strength, i.e. a weighted average of the predictions (A1111 added them): a :1 AND b :1 gives half of each, and a weight on a single prompt has no effect.
• Width and height are rounded down to a multiple of 8, as A1111 did.
• Contrast above 1 is clipped to 255 (the original overflowed the uint8 image).

Install

Copy this folder into ComfyUI/custom_nodes/ and restart ComfyUI. No dependencies beyond ComfyUI's own.

Tests

python tests/test_render.py runs the render loop with a fake sampler (no model, CPU only).

Credits

• Deforum — the extension for AUTOMATIC1111 (deforum-art/sd-webui-deforum) and the Deforum Stable Diffusion notebook, by deforum and contributors. Only code released under the MIT License was ported: the extension as of commit a7b829e (2023-08-10) and the notebook.
• Disco Diffusion — the original transform_image_3d (Maxwell Ingham, Adam Letts, Alex Spirin; MIT).
• PyTorch3D — py3d_tools.py is the extract of PyTorch3D used by Deforum (Meta Platforms; BSD).
• The 2D Perlin noise (perlin_clean.py) is our own clean-room implementation of Ken Perlin's public algorithm, matching Deforum's noise bit for bit.

The licence texts are in LICENSE.