NODES.md

Node reference

Generated from the code by tools/build_docs.py. Every node also shows this text as its description inside ComfyUI.

1 · Setup

Difforum · Setup

Difforum_Setup

Duration, framing and the grid of the model you will render with.

Say how long the clip is (seconds or frames) and how it is framed (aspect + long edge); the resolution is derived, snapped to what the target model's VAE accepts, and the frame count is snapped to the model's length grid (LTX 8k+1, MiniMax H3 17k+5 at 24 fps, Wan 4k+1). Every other Difforum node reads these params, so the whole graph agrees on one timeline.

input	type	default	notes
target	choice (feedback (SD/SDXL/Flux), LTX-2 / 2.5, MiniMax H3, Wan 2.x)	feedback (SD/SDXL/Flux)	The model you will render with. Snaps length and size to its grid: H3 17k+5 frames @ 24 fps and 32 px, LTX 8k+1 and 32 px, Wan 4k+1 and 16 px, feedback 8 px.
duration_mode	choice (seconds, frames)	seconds	Give the length in seconds or in frames.
duration	FLOAT	5.0	Clip length (seconds or frames). The target grid may round it slightly.
fps	FLOAT	24.0	Frames per second. MiniMax H3 is always 24.
aspect	choice (16:9 landscape, 9:16 vertical, 1:1 square, 4:5 social portrait, 4:3 classic, 3:4 portrait, ...)	16:9 landscape	Frame shape. 'custom' uses custom_aspect_w / h.
long_edge	INT	768	Pixels on the longer side before snapping. H3 two-stage templates render at 640 and upscale the latent 2x; the feedback look is fine at 768-1024.
seed	INT	0	Master seed for every sampler that reads Setup.
custom_aspect_w (optional)	FLOAT	16.0	Width part of a custom aspect (e.g. 2.39).
custom_aspect_h (optional)	FLOAT	9.0	Height part of a custom aspect (e.g. 1).
max_megapixels (optional)	FLOAT	0.0	Hard cap on frame area. 0 = no cap. ~0.35 on a 16 GB Mac, ~0.65 on 24 GB.

Outputs: params (DIFFORUM_PARAMS), width (INT), height (INT), frames (INT), fps (FLOAT), info (STRING)

Difforum · Workflow Switches

Difforum_Switches

Control panel for the workflow: one switch per group.

Turning a group off mutes it when it holds an output (Previz, Render) and bypasses it when it sits in the middle of the chain (Live Preview, Fill Reveal, Restyle, Look Mix, Upscale), so the rest of the graph still runs. ⌖ jumps the canvas to the group. The node runs nothing itself.

2 · Direction

Difforum · Director (timeline)

Difforum_Director

Direct the whole clip on a visual timeline.

Three tracks, edited with the mouse: Scenes (prompt + mood), Camera (pick a move, set speed / amplitude / lens / easing / audio reaction) and Energy (the denoise curve - how much each moment is re-imagined). The preview panel plays the camera back with the same engine that renders it.

One direction wire carries camera, strength, cfg and prompt travel into the Feedback Sampler or a video-model bridge. The separate outputs are there for custom graphs.

Plug pictures into images / image_1..6 and each one is pinned to a moment (the scene starts by default): the Feedback Sampler travels through them, H3 / LTX Guides anchor them, the Animatic shows them - all through the same wire. The Script button edits the whole timeline as text, one line per frame range.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: the timeline length and fps.
timeline	STRING	{"version": 2, "scenes": [{"start": 0...	Written by the timeline editor below (scenes, camera, keys, energy). Edit it there.
camera_mode	choice (2d, 3d)	2d	2d: flat moves (zoom, pan, roll). 3d: real parallax when a depth map reaches the renderer; 3D moves without depth become pseudo-3D.
look	choice (cinematic, documentary, deforum_morph, animatediff_dream, psychedelic, disco_diffusion, ...)	cinematic	One aesthetic for every renderer. Feedback Sampler: colour lock, detail, grain and extra energy. H3 / LTX: a style sentence at the head of the prompt.
transition	FLOAT	1.0	How much of each camera block eases into the next. 1 = one continuous crane move, 0 = hard cuts between blocks.
camera_scale	FLOAT	1.0	Master multiplier on every move. 0.5 = everything half as far, 2 = twice.
energy_bias	FLOAT	0.0	Moves the whole energy (denoise) curve up or down. +0.1 = more re-imagining everywhere.
variation	FLOAT	0.0	0 = exactly as drawn. Above 0, each block gets small seeded changes in speed and amount - useful for variations of the same shot.
variation_seed	INT	0	Seed for those variations: change it for another take.
clip (optional)	CLIP		Connect the CLIP of your image model to encode the scene prompts: the Feedback Sampler and Restyle then travel through them.
audio (optional)	DIFFORUM_AUDIO		From Audio Analyzer: enables the per-block audio reactions (beat pulse, bass speed...).
timeline_in (optional)	STRING		A timeline from outside: Shot Script, an LLM node, a text loader. Plain shot lines ('0s / calm / dolly_in slow / prompt'), CSV or Director JSON. The editor shows the result after each run; disconnect to edit by hand.
external (optional)	choice (replace, text only (keep drawn camera), camera only (keep drawn text), add to drawn)	replace	How timeline_in meets the drawn timeline. replace: all from outside. text only: scenes and keys from outside, your drawn camera stays. camera only: the reverse. add to drawn: both, outside wins on the same frame.
images (optional)	IMAGE		Pictures pinned to moments of the clip, in order (a batch: Keyframe Assets, Scene Stills, Batch Images). They travel on the direction wire: the Feedback Sampler passes through them, H3 / LTX Guides anchor them, the Animatic shows them.
image_1 (optional)	IMAGE		Picture 1, any size (cropped to the canvas). Pinned in order, after images.
image_2 (optional)	IMAGE		Picture 2, any size (cropped to the canvas). Pinned in order, after images.
image_3 (optional)	IMAGE		Picture 3, any size (cropped to the canvas). Pinned in order, after images.
image_4 (optional)	IMAGE		Picture 4, any size (cropped to the canvas). Pinned in order, after images.
image_5 (optional)	IMAGE		Picture 5, any size (cropped to the canvas). Pinned in order, after images.
image_6 (optional)	IMAGE		Picture 6, any size (cropped to the canvas). Pinned in order, after images.
images_at (optional)	choice (scene starts, Keys markers, spread evenly)	scene starts	Where the pictures land: the start of each scene, the Keys markers, or spread evenly from the first to the last frame.

Outputs: direction (DIFFORUM_DIRECTION), camera (DIFFORUM_CAMERA), strength (DIFFORUM_SCHEDULE), prompts (DIFFORUM_PROMPT), camera_text (STRING), info (STRING), keyframes (IMAGE), indices (STRING)

Difforum · Camera (keys)

Difforum_Camera

Keyframed camera as text, one key per line:

frame: move [speed] [intensity] [lens] [ease]

Moves blend into each other (transition 1.0 = continuous crane move, 0.0 = hard cuts). loop_mode makes the path periodic over cycle_frames so the clip loops without a crossfade. The Director timeline writes the same thing visually - use this node when you prefer typing.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
keys	STRING	0: zoom_in 0.6 0.5 40 ease_in_out 30:...	One key per line: frame: move [speed] [intensity] [lens°] [ease]. e.g. '48: orbit_right 1.2 0.8 35 ease_in_out'. Frames can be '2.5s'.
camera_mode	choice (2d, 3d)	2d	2d = flat moves; 3d = parallax with a depth map.
transition	FLOAT	1.0	1 = moves blend into each other; 0 = hard cuts at each key.
speed	FLOAT	1.0	Master speed on every key.
loop_mode	choice (harmonic, zero_mean, off)	off	harmonic / zero_mean make the path periodic over cycle_frames, so the clip loops with no crossfade (installations). off = free path.
cycle_frames	INT	0	Frames per loop. 0 = the whole clip is one cycle. Feeds Loop 'keep settled lap'.
harmonics	INT	3	harmonic loop: how many sine terms describe the path. More = closer to the keys, fewer = smoother loop.
audio (optional)	DIFFORUM_AUDIO		From Audio Analyzer: per-frame loudness, bands and beats.

Outputs: camera (DIFFORUM_CAMERA), cycle_frames (INT), info (STRING)

Difforum · Camera (expressions)

Difforum_CameraExpr

Deforum-style camera from math expressions, per axis.

0:(0), 60:(2*sin(2*pi*t/30)), with t (frame), s (seconds) and any audio band (low, beat...) when audio is connected. Each axis also has an optional schedule socket (e.g. an Audio Curve) that overrides its text. Values are per-frame increments; zoom is a per-frame scale factor.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
mode	choice (2d, 3d)	2d	2d = flat moves; 3d = parallax with a depth map.
fov	FLOAT	40.0	Lens field of view in degrees. 20-30 long lens, 40 normal, 60+ wide.
translation_x	STRING	0:(0)	Horizontal move per frame. Deforum syntax: '0:(0), 60:(2*sin(t/10))'; t = frame.
translation_y	STRING	0:(0)	Vertical move per frame, same syntax.
translation_z	STRING	0:(0)	Forward (+) / back (-) per frame. 3D only.
rotation_3d_x	STRING	0:(0)	Tilt, degrees per frame. 3D only.
rotation_3d_y	STRING	0:(0)	Pan (turn), degrees per frame. 3D only.
rotation_3d_z	STRING	0:(0.3)	Roll, degrees per frame.
zoom	STRING	0:(1.0 + 0.004sin(2pi*t/48))	Scale per frame. 1.0 = none, 1.01 = 1% closer each frame.
audio (optional)	DIFFORUM_AUDIO		Audio curves usable in expressions: amp, low, mid, high, onset, beat.
translation_x_curve (optional)	DIFFORUM_SCHEDULE		A schedule (Audio Curve, Schedule) that replaces the text for this axis.
translation_y_curve (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the text for this axis.
translation_z_curve (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the text for this axis.
rotation_3d_x_curve (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the text for this axis.
rotation_3d_y_curve (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the text for this axis.
rotation_3d_z_curve (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the text for this axis.
zoom_curve (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the text for this axis.

Outputs: camera (DIFFORUM_CAMERA), info (STRING)

Difforum · Storyboard

Difforum_Storyboard

The whole clip warped by the camera, no diffusion - in about a second.

Same engine as the Feedback Sampler with the sampler switched off, so what you see is exactly the motion the render will get: contact sheet, the frames themselves, the camera path from above, and a pacing verdict.

input	type	default	notes
init_image	IMAGE		Any still: the camera move is played over it.
every_nth	INT	8	One contact-sheet cell every N frames.
columns	INT	6	Cells per row on the contact sheet.
preview_scale	FLOAT	0.5	Working size vs. Setup. 0.5 = half size, much faster.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
depth (optional)	IMAGE		A depth map (white = near), e.g. Depth Anything V2 on the first frame. Needed for real 3D parallax; without it 3D moves fall back to pseudo-3D.
symmetry (optional)	choice (none, mirror_h, mirror_v, mirror_quad, kaleidoscope)	none	Preview an in-loop symmetry (kaleidoscope...) as the Feedback Sampler would apply it.

Outputs: sheet (IMAGE), frames (IMAGE), camera_path (IMAGE), info (STRING)

Difforum · Keyframe Images

Difforum_KeyframeImages

Pin your own pictures to moments of the clip (multikeyframing).

Place markers on the Director's Keys track (or type times: 0, 4s, 9.5s) and feed a batch of images in the same order. The keyframes then drive every renderer: the Feedback Sampler travels through them, H3 / LTX Guides anchor them, the Animatic shows them. Made for installations and experimental pieces where the image has to hit a picture on a beat.

input	type	default	notes
images	IMAGE		One picture per key, in time order (Batch Images). These pictures set the look and the content the render passes through, so make them good: same style, same resolution.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
times (optional)	STRING		Override the Director's Keys: '0, 4s, 9.5s' or frame numbers. Empty = use the Keys track; no keys = spread evenly.

Outputs: keyframes (IMAGE), indices (STRING), info (STRING)

Difforum · Animatic (previz)

Difforum_Animatic

Previz the whole shot before rendering anything heavy.

Plays the Director's camera over your first frame (or a stand-in plate) at low resolution, in about a second, with the timecode, the active scene prompt, the camera move, the energy level and the keyframe markers burnt in. Image keyframes, when connected, show up at their moments. Send it to Create Video + Save Video; the Director's "Previz only" button mutes the render outputs so only this runs.

input	type	default	notes
direction	DIFFORUM_DIRECTION		The Director to previz.
preview_scale	FLOAT	0.4	Size vs. Setup. 0.4 = small and instant; 1.0 = full size.
overlay	BOOLEAN	True	Burn in timecode, camera move, scene prompt, energy bar and key markers.
init_image (optional)	IMAGE		First frame to move over. Empty = a grid plate.
depth (optional)	IMAGE		Depth of the first frame, for real parallax in 3d mode.
key_images (optional)	IMAGE		Keyframe Images: shown at their moments.
key_indices (optional)	STRING		Their frame numbers.

Outputs: frames (IMAGE), fps (FLOAT), info (STRING)

Difforum · Shot Script (timeline from text)

Difforum_ShotScript

Write the Director timeline from text instead of drawing it.

One beat per line: TIME | MOOD | CAMERA | PROMPT (0s | calm | dolly_in slow | misty forest), 4s | key: the light breaks for a key marker, or just prompts without times (spread evenly). A CSV with a header row (time,mood,camera,prompt,key,energy) or a Director JSON works too.

The text can come from the box, from a .txt / .csv / .json file in ComfyUI's input folder (re-read when it changes; examples/01_infinite_zoom.txt and the other shipped scripts work too), or from any node that outputs a STRING - an LLM node, a text loader, a spreadsheet export - into script_in. Wire timeline into the Director's timeline_in. llm_instructions is a ready prompt that asks an LLM for this format.

input	type	default	notes
script	STRING	# TIME	MOOD
file	STRING		A .txt / .csv / .json in ComfyUI/input (e.g. shots/scene01.txt), or a shipped example: examples/01_infinite_zoom.txt. Re-read whenever it changes. Overrides the box.
params (optional)	DIFFORUM_PARAMS		From Setup: converts seconds to frames and flags beats past the end.
script_in (optional)	STRING		Text from another node (an LLM, a text file loader, a spreadsheet export). Overrides the box and the file. Feed the llm_instructions output to the LLM as its prompt.

Outputs: timeline (STRING), llm_instructions (STRING), info (STRING)

Difforum · Keyframe Assets (folder)

Difforum_KeyframeAssets

Your own styled pictures as the keyframes - no look pass to render.

Point folder at a folder inside ComfyUI/input (stills made in any image model, Photoshop, a photo shoot...). Each picture lands on a moment:

• Director keys: in order, on the Keys track markers.
• filename: the name says when - 0s_wide.png, 4.5s_close.png, f096.png, or a leading frame number 0096_rise.png.
• spread evenly: first at frame 0, last at the end.

fit matches the canvas: cover crops, contain pads with gray (and the masks output tells Fill Reveal what to paint), stretch distorts. Wire keyframes + indices into H3 Guides / LTX Guides / Animatic, and first / last into a first-last-frame model. An images batch from other nodes can replace the folder.

input	type	default	notes
folder	STRING	difforum_keys	A folder inside ComfyUI/input holding the stills (png / jpg / webp), sorted by name.
timing	choice (Director keys, filename, spread evenly)	filename	Director keys: the pictures land on the Keys markers in order. filename: '0s_x.png', '4.5s_x.png', 'f096.png' or '0096_x.png'. spread evenly: first to last over the clip.
fit	choice (cover (crop), contain (pad gray), stretch)	cover (crop)	cover crops to the canvas, contain pads with gray (the masks output lets Fill Reveal paint the bars), stretch distorts.
direction (optional)	DIFFORUM_DIRECTION		Gives the Keys track and the canvas size.
params (optional)	DIFFORUM_PARAMS		Canvas size and fps when no direction is connected.
images (optional)	IMAGE		A batch from other nodes instead of the folder (one per key, in order).
times (optional)	STRING		Override: '0, 4s, 9.5s' or frame numbers.

Outputs: keyframes (IMAGE), indices (STRING), first (IMAGE), last (IMAGE), masks (MASK), info (STRING)

Difforum · Scene Stills (storyboard)

Difforum_SceneStills

One still per scene of the timeline: the storyboard, painted by the image model.

Each scene prompt (with your style in front) becomes a picture at the frame where the scene starts. continuity paints every still over the one before it, so palette, light and layout carry from beat to beat instead of jumping; 0 makes each one from scratch. Feed keyframes + indices to H3 Guides / LTX Guides and a video model animates between pictures that already tell the story: a handful of images instead of a look pass.

Stills are made at long_edge (the image model's own size) in the canvas aspect, so they stay sharp when the video model renders smaller.

input	type	default	notes
direction	DIFFORUM_DIRECTION		The Director timeline: one still per scene that has a prompt.
model	MODEL		An image model (SDXL, SD1.5, Flux...).
clip	CLIP
vae	VAE		The VAE that matches the model.
style	STRING		The look shared by every still, written once: medium, light, lens, palette. It goes in front of each scene prompt.
negative	STRING	blurry, low quality, watermark, text,...	What no still should show.
steps	INT	24	Sampling steps per still. 24 for a base model, 4-8 with a turbo / DMD2 model.
cfg	FLOAT	5.5	Prompt strength. 5-6 for a base model, 1-2 for turbo / DMD2.
sampler_name	choice (euler, lcm)	euler	Sampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo.
scheduler	choice (normal, sgm_uniform)	normal	Noise schedule. karras / normal for regular models, sgm_uniform for turbo.
seed	INT	7	Same seed for every still, so they share composition habits and texture.
continuity	FLOAT	0.5	How much each still is painted over the one before. 0 = every still from scratch (free composition, the look can jump). 0.5 = palette and layout carry over. 1 = small changes only.
long_edge	INT	1024	Size the stills are made at, in the canvas aspect. Use the image model's own size (1024 SDXL, 768 SD1.5): the video model scales them down itself.
first_image (optional)	IMAGE		Your own picture as the first still; the rest are painted after it.

Outputs: keyframes (IMAGE), indices (STRING), first (IMAGE), last (IMAGE), info (STRING)

Difforum · Travel Conditioning (prompt per frame)

Difforum_TravelConditioning

The Director's prompt travel as one CONDITIONING with a prompt per frame.

Samplers that render a whole clip at once - AnimateDiff, or any batch KSampler - take a single conditioning. This node stacks the timeline's scene prompts, blended from scene to scene, into a batch as long as the clip, so frame 40 is sampled with the prompt of frame 40: the Deforum prompt schedule, on AnimateDiff. Connect a CLIP to the Director (it encodes the scenes). The batch is as long as the Director's clip, or as the images / latent you connect (a rendered video of another length).

input	type	default	notes
direction (optional)	DIFFORUM_DIRECTION		A Director with a CLIP connected: its scene prompts become one prompt per frame.
prompts (optional)	DIFFORUM_PROMPT		Or the output of a Prompt Travel node.
images (optional)	IMAGE		Optional: the frames that will be sampled; the travel is stretched to their count.
latent (optional)	LATENT		Optional: the latent batch that will be sampled; the travel is stretched to its count.

Outputs: positive (CONDITIONING), frames (INT), info (STRING)

3 · Curves & Prompts

Difforum · Schedule

Difforum_Schedule

A per-frame curve from Deforum keyframe syntax: 0:(0.4), 48:(0.6+0.1*sin(t/8)).

Variables: t/f (frame), s (seconds), fps, max_f, and every audio band when an Audio Analyzer is connected. Easing applies between keyframes.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
schedule	STRING	0:(0.45), 60:(0.6), 119:(0.45)	Values over time: 'frame:(value)', e.g. '0:(0.45), 60:(0.6)'. Frames can be '2s'. Expressions allowed: t = frame, amp/low/beat with audio.
easing	choice (linear, ease_in, ease_out, ease_in_out, step)	ease_in_out	How values move between keys.
audio (optional)	DIFFORUM_AUDIO		From Audio Analyzer: per-frame loudness, bands and beats.

Outputs: schedule (DIFFORUM_SCHEDULE), info (STRING)

Difforum · Schedule Plot

Difforum_SchedulePlot

Draw any schedule as an image (and a one-line summary).

input	type	default	notes
schedule	DIFFORUM_SCHEDULE		Any schedule or curve to draw.
width	INT	512	Plot width in pixels.
height	INT	200	Plot height in pixels.

Outputs: plot (IMAGE), info (STRING)

Difforum · Audio Analyzer

Difforum_AudioAnalyzer

Audio -> per-frame bands: amp, low, mid, high, onset, beat (0..1).

Use the bands in any expression (1 + 0.05*low), in a Director camera block's audio reaction, or turn one into a curve with Audio Curve.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
audio	AUDIO		Your track (Load Audio).
smoothing	FLOAT	0.25	0 = every transient, 0.5+ = slow envelope. 0.2-0.3 reads well on camera.
beat_sensitivity	FLOAT	1.5	Higher = fewer, stronger beats detected.
offset_seconds	FLOAT	0.0	Start reading the track here, to sync with an edit.

Outputs: audio_curves (DIFFORUM_AUDIO), bands_plot (IMAGE), info (STRING)

Difforum · Audio Curve

Difforum_AudioCurve

One audio band -> a ready curve: e.g. bass-pumped energy (low, add, base 0.45, amount 0.2) or a beat-pulsed cfg.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
audio_curves	DIFFORUM_AUDIO		From Audio Analyzer.
source	choice (amp, low, mid, high, onset, beat)	low	Which band drives the curve: amp (loudness), low / mid / high bands, onset (attacks), beat (pulses).
combine	choice (add, subtract, multiply)	add	How the audio is applied to base: add, subtract or multiply.
base	FLOAT	0.45	Value when the band is silent (e.g. 0.45 denoise).
amount	FLOAT	0.2	How far the band pushes the value at its loudest.
smoothing	FLOAT	0.2	Extra smoothing of the band before it is used.

Outputs: schedule (DIFFORUM_SCHEDULE)

Difforum · Prompt Travel

Difforum_PromptTravel

Prompt travel from frame: prompt lines (or 2.5s: prompt), blended per frame. Encodes each prompt once and blends lazily, so long clips on big text encoders stay light. build_batched adds one batched conditioning (frame i = prompt i) for batch samplers such as AnimateDiff.

input	type	default	notes
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
clip	CLIP		CLIP of the image model.
prompts	STRING	0: a serene misty forest, soft light ...	One prompt per line: 'frame: text' or '2.5s: text'. Consecutive prompts blend.
easing	choice (linear, ease_in, ease_out, ease_in_out, step)	ease_in_out	How each prompt blends into the next.
build_batched (optional)	BOOLEAN	False	Also build the batched output (costs RAM on long clips).

Outputs: prompts (DIFFORUM_PROMPT), first_frame_cond (CONDITIONING), batched (CONDITIONING), info (STRING)

4 · Render

Difforum · Feedback Sampler

Difforum_FeedbackSampler

The Deforum look on any image model: every frame is the last one, moved by the camera and re-imagined by the sampler.

Plug a Director direction wire and a first frame; that is the whole setup. Works with SD1.5, SDXL, Flux, SD3.5 and turbo/LCM/DMD2 distills (plain MODEL / VAE / CONDITIONING). Connect a depth map (e.g. Depth Anything V2 on the first frame) and set the Director to 3d for real parallax - the depth then follows the image frame by frame.

Outputs the frames, the tracked depth per frame (for comp / relight / video-model guides) and a run report.

input	type	default	notes
model	MODEL		The image model (SD1.5 / SDXL / Flux / turbo). MODEL output of your checkpoint or LoRA chain.
positive	CONDITIONING		Look / style prompt. Used for every frame when no scene prompts reach the sampler.
negative	CONDITIONING		Negative conditioning (CLIP Text Encode).
vae	VAE		The VAE that matches the model.
init_image	IMAGE		The first frame. Everything grows from this image.
steps	INT	20	Steps for a full denoise. With step scaling (default) a frame at energy 0.5 runs half of them, like Deforum.
cfg	FLOAT	6.0	Prompt strength. Turbo / LCM / DMD2 models want 1-2; regular SDXL 4-7.
sampler_name	choice (euler, lcm)	euler	Sampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo.
scheduler	choice (normal, sgm_uniform)	normal	Noise schedule. karras / normal for regular models, sgm_uniform for turbo.
cadence	INT	1	Diffuse every Nth frame; the frames in between ride the camera and crossfade. 2 is almost invisible and twice as fast.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
options (optional)	DIFFORUM_OPTIONS		Difforum · Render Options, for manual control.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
strength (optional)	DIFFORUM_SCHEDULE		A schedule that replaces the Director's energy curve (denoise per frame).
cfg_curve (optional)	DIFFORUM_SCHEDULE		A schedule for cfg per frame.
prompts (optional)	DIFFORUM_PROMPT		Prompt travel from Prompt Travel (or the Director with CLIP connected).
depth (optional)	IMAGE		Depth of the first frame (white = near), or a per-frame batch.
control_net (optional)	CONTROL_NET		Optional ControlNet (depth, canny...) applied on every diffused frame.
control_image (optional)	IMAGE		Per-frame control images; empty = the frame itself.
control_strength (optional)	FLOAT	0.6	ControlNet strength.
energy (optional)	FLOAT	0.5	Denoise per frame when no Director or strength curve is connected. It is how much each frame is re-imagined: 0.3 = steady, 0.5 = classic Deforum morph, 0.7+ = the image changes every few frames.
key_images (optional)	IMAGE		Keyframe Images: the travel steers toward each picture and lands on it at its frame. This is multikeyframing.
key_indices (optional)	STRING		Frame numbers of the key pictures (Keyframe Images 'indices').
key_pull (optional)	FLOAT	0.65	How strongly the frame is pulled toward the next key picture before diffusion. 1 = arrives exactly on it, 0.6 = strong resemblance, 0.3 = a hint.
key_approach (optional)	INT	12	Frames before a key over which the pull ramps up. Short (6-12) = sudden arrival, long (24-48) = the image slowly becomes the key. Long shots want long approaches.

Outputs: frames (IMAGE), depth (IMAGE), report (STRING)

Difforum · Live Sampler

Difforum_LiveSampler

Realtime Difforum: queue once and watch it play inside the node.

Pair with a 1-4 step model (SD-Turbo, SDXL-Turbo, LCM, DMD2) at ~512 px. live_source turns it into a magic mirror (webcam "0" or a video path); stream_dir / spout_name feed OBS, Resolume or TouchDesigner. Only the last keep_frames frames are returned, so long sessions do not fill RAM.

input	type	default	notes
model	MODEL		The image model (SD1.5 / SDXL / Flux / turbo). MODEL output of your checkpoint or LoRA chain.
positive	CONDITIONING		Positive conditioning (CLIP Text Encode).
negative	CONDITIONING		Negative conditioning (CLIP Text Encode).
vae	VAE		The VAE that matches the model.
init_image	IMAGE		The first frame. Everything grows from this image.
run_frames	INT	480	How many frames to run (the loop wraps the camera).
steps	INT	2	Steps per frame. Turbo models: 1-4.
cfg	FLOAT	1.2	Prompt strength. Turbo / LCM / DMD2 models want 1-2; regular SDXL 4-7.
sampler_name	choice (euler, lcm)	lcm	Sampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo.
scheduler	choice (normal, sgm_uniform)	sgm_uniform	Noise schedule. karras / normal for regular models, sgm_uniform for turbo.
cadence	INT	1	Diffuse every Nth frame for speed.
target_fps	FLOAT	0.0	Cap the speed. 0 = as fast as possible.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
options (optional)	DIFFORUM_OPTIONS		Difforum · Render Options: manual colour, detail, symmetry, depth and chunk settings. Without it the Director's look decides.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
strength (optional)	DIFFORUM_SCHEDULE		A schedule for denoise per frame; replaces the Director's energy curve.
prompts (optional)	DIFFORUM_PROMPT		Prompt travel from Prompt Travel (or the Director with CLIP connected).
depth (optional)	IMAGE		A depth map (white = near), e.g. Depth Anything V2 on the first frame. Needed for real 3D parallax; without it 3D moves fall back to pseudo-3D.
energy (optional)	FLOAT	0.5	Denoise per frame when no Director is connected.
loop_camera (optional)	BOOLEAN	True	Repeat the camera when the run is longer than the clip.
live_source (optional)	STRING		'' = off, '0' = webcam, or a video path: mixed in before each frame.
source_blend (optional)	FLOAT	0.9	How much of the live source enters each frame.
stream_dir (optional)	STRING		Folder to write live PNG frames to (for another app).
spout_name (optional)	STRING		Spout sender name (Windows) for Resolume / TouchDesigner / OBS.
keep_frames (optional)	INT	240	Frames kept in memory for the output (ring buffer).
live_preview (optional)	BOOLEAN	True	Show the frames inside the node while running.

Outputs: frames (IMAGE), report (STRING)

Difforum · Render Options

Difforum_RenderOptions

Everything the samplers can fine-tune, kept off the sampler itself.

Without this node the look (colour lock, sharpen, grain) comes from the Director's look and the rest uses sensible defaults. Connect it to take manual control: colour anchoring, detail guard, in-loop symmetry, 3D depth calibration, noise seeding and chunked rendering.

input	type	default	notes
color_coherence	FLOAT	0.8	How hard colours are held to the anchor.
color_mode	choice (lab, rgb, none)	lab	How colour is matched to the anchor: lab (perceptual), rgb, or none.
anchor_mode	choice (scene, first, rolling, none)	scene	scene = re-anchor colour at each prompt scene; first = hold frame 0; rolling = previous key; none = free.
sharpen	FLOAT	0.2	Sharpening after each diffused frame; keeps feedback from going soft.
noise	FLOAT	0.02	Grain added before each frame is diffused; feeds new detail.
hole_noise	FLOAT	0.25	Extra noise where the camera reveals new area, so it is repainted.
symmetry	choice (none, mirror_h, mirror_v, mirror_quad, kaleidoscope)	none	Symmetry applied inside the loop (kaleidoscope for mandalas).
symmetry_segments	INT	6	Kaleidoscope segments.
border	choice (reflection, border, zeros)	reflection	What fills the edges the camera exposes: reflection (mirror), border (stretch), zeros.
depth_tracking	choice (follow, static)	follow	follow = depth is re-projected with the image every frame.
translation_scale	FLOAT	1.0	3D motion strength vs. the depth map's range.
near	FLOAT	1.0	Depth calibration: distance of the nearest depth value.
far	FLOAT	100.0	Depth calibration: distance of the farthest depth value.
invert_depth	BOOLEAN	False	Enable if near things are dark in your depth map.
step_scaling	choice (by energy (fast), fixed)	by energy (fast)	by energy: a frame at denoise 0.5 runs half the steps, like Deforum (about 2x faster at the same look). fixed: every frame runs all steps.
seed_mode	choice (fixed, increment)	fixed	fixed = same noise every frame (calmer texture).
start_frame	INT	0	Feedback Sampler: render a chunk; init_image is the frame at start_frame.
end_frame	INT	0	0 = to the end.

Outputs: options (DIFFORUM_OPTIONS)

Difforum · Restyle (Deforum / AnimateDiff look)

Difforum_Restyle

Give a video-model render the Deforum / AnimateDiff / Disco look.

MiniMax H3 or LTX supplies the motion; an image model (SDXL, SD1.5, Flux, turbo distills) re-paints every frame in your look, with the previous stylized frame carried along the video's optical flow and mixed in. That feedback is what made Deforum morph and AnimateDiff boil - here it rides on the video model's motion instead of a synthetic camera.

Styles: clean restyle (steady), animatediff boil (texture re-rolls each frame), deforum morph (strong feedback smear), disco flicker (high denoise, per-frame seed, loose colour). custom uses feedback / seed_mode / color_hold as set. Connect a Director to use its scene prompts and energy.

input	type	default	notes
video	IMAGE		Frames from H3 / LTX / any video (Get Video Components).
model	MODEL		An image model in the look you want. Turbo SDXL / DMD2 at 4-6 steps is fast and good.
positive	CONDITIONING		The look, e.g. 'oil painting, thick brush strokes'.
negative	CONDITIONING		Negative conditioning (CLIP Text Encode).
vae	VAE		The VAE that matches the model.
style	choice (clean restyle, animatediff boil, deforum morph, disco flicker, custom)	deforum morph	deforum morph: strong feedback, smears along motion. animatediff boil: texture re-rolls each frame. disco flicker: high energy, loose colour. clean restyle: steady painted version. custom: uses feedback / seed_mode / color_hold below.
denoise	FLOAT	0.45	How much each frame is re-painted. 0.3 keeps the render, 0.6+ re-imagines it.
steps	INT	12	Full-denoise steps; each frame runs steps x denoise.
cfg	FLOAT	5.0	Prompt strength. Turbo models: 1-2.
sampler_name	choice (euler, lcm)	euler	Sampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo.
scheduler	choice (normal, sgm_uniform)	normal	Noise schedule. karras / normal for regular models, sgm_uniform for turbo.
cadence	INT	1	Re-paint every Nth frame; the rest follow the flow. 2 = twice as fast.
long_edge	INT	1024	Working size. 0 = the video's size. Upscale afterwards for 2K.
seed	INT	0	Noise seed.
direction (optional)	DIFFORUM_DIRECTION		Director: scene prompts (with CLIP) and the energy curve.
prompts (optional)	DIFFORUM_PROMPT		Prompt travel; otherwise the Director's scene prompts, otherwise positive.
feedback (optional)	FLOAT	0.35	custom: share of the previous stylized frame in the next one.
seed_mode (optional)	choice (fixed, per frame)	fixed	custom: per frame = texture re-rolls every frame (boil / flicker).
color_hold (optional)	FLOAT	0.5	custom: how much each frame keeps the source colours.
follow_energy (optional)	BOOLEAN	True	With a Director: denoise follows its energy curve (0.5 = as set).
flow_scale (optional)	FLOAT	0.5	Resolution of the motion estimate that carries the previous frame. 0.5 = fast.
control_net (optional)	CONTROL_NET		A ControlNet that holds the clip's structure while the look is re-painted, so denoise can go up to 0.6-0.8 (the real Deforum / AnimateDiff range) without losing the shapes. Depth or canny with control_image; tile with the clip itself.
control_image (optional)	IMAGE		Per-frame control maps of the clip (Depth Anything 3, Canny...). Empty = the clip frames themselves (tile / union ControlNets).
control_strength (optional)	FLOAT	0.6	How hard the structure is held. 0.4 loose, 0.6 balanced, 0.9 locked.

Outputs: frames (IMAGE), report (STRING)

5 · Video model bridges

Difforum · Guide Frames

Difforum_GuideFrames

Warp one anchor image along the Director's camera path.

Produces a guide video whose motion is exactly the camera you drew, for video models that take control/guide frames (LTX guides, Wan VACE, H3 first/last frame). Revealed areas are filled with neutral gray and marked in the mask, using the convention the target expects.

input	type	default	notes
anchor_image	IMAGE		The image the camera travels over (usually your first frame).
mask_convention	choice (1 = generate (VACE / LTX / H3), 1 = keep)	1 = generate (VACE / LTX / H3)	Which value the mask uses for 'generate': 1 = generate (VACE / LTX / H3 / Fill Reveal) or 1 = keep.
hole_fill	choice (gray, stretch edge, black)	gray	What goes into the area the camera uncovers before AI fills it: gray (neutral, best for Fill Reveal and H3), stretched edge, or black.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
depth (optional)	IMAGE		Depth of the anchor, for real parallax in 3d mode.
translation_scale (optional)	FLOAT	1.0	Strength of 3D translation vs. the depth range.

Outputs: guide_frames (IMAGE), masks (MASK), info (STRING)

Difforum · Keyframes

Difforum_Keyframes

Pick keyframes out of any frame batch (a Storyboard, a Feedback render, guide frames) for a video model, on that model's frame grid.

Outputs the keyframes, their indices, a full-length sparse batch (black except at keyframes - what LTXVAddGuidesFromBatch expects), the first and last frame (for first-last-frame models) and the snapped clip length.

input	type	default	notes
frames	IMAGE		The clip to pick keyframes from (Guide Frames, Storyboard, a Feedback render...). Keyframes carry its look into H3 / LTX: better frames here = better video.
grid	choice (LTX-2 / 2.5 (8k+1), MiniMax H3 (17k+5), Wan 2.x (4k+1), any)	LTX-2 / 2.5 (8k+1)	Snap keyframes to the video model's latent grid (H3 17k+5, LTX 8k+1, Wan 4k+1) so each one lands on a real latent frame.
every_seconds	FLOAT	1.0	How many keyframes: one every N seconds. 1 = many, strong guidance; 2 = fewer, more freedom; 0 = first and last only. H3 Guides then keeps up to max_guides.
fps	FLOAT	24.0	Frames per second of the clip (from Setup).
indices (optional)	STRING		Exact frames instead of spacing, e.g. '0, 48, 96, -1' (-1 = last).
masks (optional)	MASK		Guide Frames masks, so Fill Reveal can repaint what the camera uncovered.

Outputs: keyframes (IMAGE), indices (STRING), sparse_batch (IMAGE), first_frame (IMAGE), last_frame (IMAGE), length (INT), key_masks (MASK), info (STRING)

Difforum · Camera → Prompt

Difforum_CameraPrompt

The camera, in words - for prompt-driven video models (MiniMax H3, LTX, Seedance, Veo...). Uses the Director's blocks when available, or reads any camera track and names its moves.

input	type	default	notes
format	choice (sentence, timed lines, prompt suffix, H3 structured)	sentence	sentence: one line of camera direction (+ look) to append to any prompt. timed lines: [0.0s-2.5s] camera ... per block. prompt suffix: 'Camera: ...'. H3 structured: the whole Director timeline (scenes, camera, keys, look) written in MiniMax H3's native prompt format.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
prefix (optional)	STRING		What is on screen. sentence / suffix: put before the camera text. H3 structured: the opening composition ([Shot 1]); in reference mode, what <Picture 1> shows.
include_look (optional)	BOOLEAN	True	Append the Director's look sentence (deforum morph, stop-motion...).
h3_mode (optional)	choice (T2VA (text only), I2VA (first frame), FL2VA (first + last frame), reference (ref2va / guides))	reference (ref2va / guides)	H3 structured only. Which H3 task the prompt is for: sets the alignment line (I2VA / FL2VA) or the six full-reference sections (ref2va with H3 Guides, <Picture 1> as the first frame).
soundscape (optional)	STRING		H3 structured: overall_soundscape - ambience and physical sounds only (wind, footsteps, rain). Empty = soft natural ambience.
music (optional)	STRING		H3 structured: non_diegetic_music - instruments, tempo, dynamics, no mood words. Empty = N/A (no score).
cuts (optional)	BOOLEAN	False	H3 structured: off = one continuous take (scene changes become transformations). On = every scene starts a new [Shot N] at its cut time.

Outputs: text (STRING)

Difforum · LTX Guides

Difforum_LTXGuides

Put Difforum keyframes into an LTX-2 / 2.5 latent as guides.

Wraps ComfyUI's core LTXVAddGuide once per keyframe, so the camera you drew becomes the LTX shot. Feed keyframes + indices from the Keyframes node (from a Storyboard, Guide Frames or a Feedback render), and an empty LTX latent whose length matches (Setup target = LTX).

input	type	default	notes
positive	CONDITIONING		LTX positive conditioning.
negative	CONDITIONING		LTX negative conditioning.
vae	VAE		LTX video VAE.
latent	LATENT		Empty LTX latent (length and size from Keyframes / Setup).
strength	FLOAT	0.7	Guide strength of the keyframes in between. 0.6-0.8 = follows the path, 1 = locked.
first_strength	FLOAT	1.0	Frame 0 usually locks the look: keep it high.
last_strength	FLOAT	0.7	Strength of the last keyframe. High = the shot must end on it.
keyframes (optional)	IMAGE		Keyframe pictures (from Keyframes or Keyframe Images).
indices (optional)	STRING	0	Frame number of each keyframe, comma separated, e.g. 0,48,96,123.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.

Outputs: positive (CONDITIONING), negative (CONDITIONING), latent (LATENT), info (STRING)

Difforum · H3 Shot

Difforum_H3Shot

Everything a MiniMax H3 first-last-frame shot needs, from Difforum.

Give it the frames of a Storyboard / Feedback render / Guide Frames and it returns the first and last frame of the chosen segment, the length on the H3 17k+5 grid, the size on the 32 px grid, and the camera move of that segment in words for the prompt. Clips longer than one H3 generation are split into segments: render segment 0, then feed its last frame as the next segment's first frame.

input	type	default	notes
frames	IMAGE		The clip whose first and last frame H3 should connect (Guide Frames, Storyboard, Feedback render).
segment_length	choice (auto (from frames), 124 (~5s), 243 (~10s), 362 (~15s), 481 (~20s))	auto (from frames)	Frames per H3 generation. Clips longer than this are split into segments.
segment	INT	0	Which segment to render: 0, then 1 with segment 0's last frame as its first frame...
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
shot_description (optional)	STRING		What is on screen at the start: subject, place, light. Goes into the prompt.
masks (optional)	MASK		Guide Frames masks: returns the last frame's revealed area for Fill Reveal.
include_look (optional)	BOOLEAN	True	Add the Director's look as the style of the shot.
prompt_style (optional)	choice (H3 structured, sentence)	H3 structured	H3 structured: this segment's scenes, camera, keys and look in MiniMax H3's native FL2VA format (alignment line + fields). sentence: description + camera sentence + look, as plain text.
soundscape (optional)	STRING		overall_soundscape: ambience and physical sounds only.
music (optional)	STRING		non_diegetic_music: instruments, tempo, dynamics. Empty = N/A.

Outputs: first_frame (IMAGE), last_frame (IMAGE), length (INT), width (INT), height (INT), prompt (STRING), segments (INT), last_mask (MASK), info (STRING)

Difforum · H3 Guides

Difforum_H3Guides

Anchor Difforum keyframes inside a MiniMax H3 generation.

Wraps ComfyUI's core MiniMaxH3AddGuide once per keyframe, so the camera you drew becomes the H3 shot: feed keyframes + indices from the Keyframes node (grid H3) and the positive + AV latent from MiniMax H3 Reference to Video (or Image to Video). Optionally anchor a soundtrack at frame 0, so an audio-reactive direction and H3's own audio stay in sync.

H3 is trained with a few guides per clip: max_guides keeps the first, the last and evenly spaced ones in between.

input	type	default	notes
positive	CONDITIONING		Positive conditioning from MiniMax H3 Reference to Video.
latent	LATENT		The MiniMax H3 AV latent.
vae	VAE		MiniMax H3 video VAE.
max_guides	INT	4	Most keyframes anchored inside the generation. The first and last are kept, the rest evenly spaced. More = H3 follows your frames closely; fewer = more of H3's own motion and invention. 4 for a camera path, 6-8 for a look pass.
skip_first	BOOLEAN	False	Enable when frame 0 is already set (e.g. Image to Video first_frame).
keyframes (optional)	IMAGE		Keyframe pictures (Keyframes, Keyframe Images or Fill Reveal).
indices (optional)	STRING	0	Frame number of each keyframe, comma separated, e.g. 0,48,96,123.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
audio_vae (optional)	VAE		MiniMax H3 audio VAE, needed with audio.
audio (optional)	AUDIO		Soundtrack anchored at frame 0.

Outputs: positive (CONDITIONING), info (STRING)

Difforum · H3 Refine Guides

Difforum_H3RefineGuides

Make H3 guides fit a two-stage refine.

MiniMax H3 guides (first / last frame, Add Guide, Difforum H3 Guides) are encoded at the size of the first render and shared with the target's spatial grid, so the refine pass after a latent upscale fails with a shape mismatch if it reuses the same conditioning. This node re-encodes every guide at the upscaled latent's size: from the original full-size pixels when Difforum H3 Guides added it, otherwise by decoding the small guide, resizing it and encoding it again. drop guides removes the video guides instead and lets the refine follow the upscaled render alone. References (ref2va pictures) keep their own size and are left untouched.

input	type	default	notes
positive	CONDITIONING		The same conditioning the first H3 pass used (with its guides / first-last frames).
latent	LATENT		The upscaled AV latent (after the H3 Latent Upscaler), so the guides match its size.
vae	VAE		MiniMax H3 video VAE.
mode	choice (re-encode, drop guides)	re-encode	re-encode: guides rebuilt at the refine size (from the original pixels when Difforum H3 Guides added them). drop guides: the refine follows the upscaled render alone.

Outputs: positive (CONDITIONING), info (STRING)

Difforum · Fill Reveal (AI)

Difforum_FillReveal

Complete what the camera reveals, with an image model (AI hole fill).

Guide Frames, H3 Shot and Keyframes mark the area the camera uncovers (outside the original picture) in a mask. This node repaints only that area by inpainting, so edges become new, coherent scenery instead of gray or stretched pixels. The known pixels are kept exactly.

Works with any image model through ComfyUI's core InpaintModelConditioning: a dedicated inpaint model (SDXL inpainting, Flux Fill) gives the cleanest seams, a regular checkpoint works too. Run it only on the frames a video model will see (first / last frame, keyframes) - it is one diffusion per frame.

input	type	default	notes
images	IMAGE		Frames with holes where the camera uncovered new area (gray from Guide Frames).
masks	MASK		1 = area to fill (Guide Frames default convention).
model	MODEL		An image model; an inpainting checkpoint (SDXL inpainting, Flux Fill) gives the cleanest seams.
positive	CONDITIONING		Describe the scene so the fill matches it.
negative	CONDITIONING		Negative conditioning (CLIP Text Encode).
vae	VAE		The VAE that matches the model.
frames	choice (all, first, last, first + last)	all	Which frames to paint: all keyframes, or only the first / last (FL2VA).
steps	INT	24	Inpainting steps.
cfg	FLOAT	5.0	Prompt strength. Turbo / LCM / DMD2 models want 1-2; regular SDXL 4-7.
sampler_name	choice (euler)	euler	Sampler. euler / dpmpp_2m for regular models, euler_ancestral or lcm for turbo.
scheduler	choice (normal)	normal	Noise schedule. karras / normal for regular models, sgm_uniform for turbo.
grow	INT	16	Pixels the mask is grown into the known image, to hide the seam.
feather	INT	12	Soft edge, in pixels, where the painted area meets the original.
seed	INT	0	Fill seed; change it for another fill.

Outputs: images (IMAGE), info (STRING)

6 · Export

Difforum · Camera Export (AE / Blender / JSON)

Difforum_CameraExport

Send the Difforum camera to After Effects, Blender or any tool (JSON).

Writes to output/difforum/:

• .jsx - run in After Effects: a 3D camera (3d mode) or a transform null that reproduces the 2D move exactly (2d mode).
• .py - run in Blender's Text Editor: an animated camera, fps, resolution and frame range matched.
• .json - camera-to-world matrices + focal length per frame. Composite titles, 3D elements or a relight pass on top of the AI render with the exact same camera.

input	type	default	notes
formats	choice (all, after effects, blender, json)	all	Which files to write to output/difforum: After Effects .jsx, Blender .py, JSON.
filename_prefix	STRING	difforum_camera	File name prefix.
blender_unit_scale	FLOAT	0.1	Metres per Difforum scene unit.
direction (optional)	DIFFORUM_DIRECTION		The Director's wire: camera, energy curve, scene prompts, keys and look in one cable.
params (optional)	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
camera (optional)	DIFFORUM_CAMERA		A camera track from Camera (keys), Camera (expressions) or Camera Import. Overrides the Director's camera.
translation_scale (optional)	FLOAT	1.0	Match the sampler's value so the exported move matches the render.

Outputs: paths (STRING), json (STRING)

Difforum · Camera Import

Difforum_CameraImport

Bring a camera in from Blender, After Effects or JSON.

Accepts the files written by Camera Export and by the helper scripts in tools/ (export the active Blender camera, or the selected After Effects camera). The result drives the Feedback Sampler, Storyboard or Guide Frames like any Difforum camera, so a move blocked in 3D becomes the AI shot.

input	type	default	notes
file	choice ((put a camera .json in ComfyUI/input))		A camera .json in ComfyUI/input: from Camera Export, the Blender exporter or the After Effects script in tools/.
params	DIFFORUM_PARAMS		From Setup: frame count, fps, width and height. Plug the same Setup everywhere.
retime	choice (match frames, keep source length)	match frames	match frames = stretch the move to Setup's length; keep source length = frame for frame.
json_text (optional)	STRING		Paste JSON here instead of choosing a file.

Outputs: camera (DIFFORUM_CAMERA), info (STRING)

7 · Post

Difforum · Loop

Difforum_Loop

Make a clip loop.

• keep settled lap - for renders made with a looping camera (Director / Camera loop_mode) over several laps: keeps the last lap, no blending.
• flow crossfade - morphs the tail into the head along optical flow.
• ping-pong - forward then backward (motion reverses). Reports how visible the seam is.

input	type	default	notes
frames	IMAGE		A batch of frames (a clip).
method	choice (keep settled lap, flow crossfade, ping-pong)	flow crossfade	keep settled lap: render several laps, keep the last (perfect loop). flow crossfade: blend the end into the start along motion. ping-pong: forward then back.
cycle_frames	INT	0	keep settled lap: frames per lap (the Camera node outputs it).
blend_frames	INT	12	flow crossfade: overlap length.

Outputs: frames (IMAGE), seam_info (STRING)

Difforum · Look Mix

Difforum_LookMix

Keep the experimental look while a video model supplies the motion.

Video models such as MiniMax H3 render fluid, coherent motion - and smooth away the boiling, flickering texture that makes Deforum, AnimateDiff and Disco Diffusion feel alive. Feed the video-model render and the Feedback Sampler look pass (same length or not; it is resampled in time):

• detail transfer - H3's motion, the look pass's high-frequency texture
• colour + detail - also pulls each frame's palette to the look pass
• flicker cuts - every Nth frame (seeded jitter) swaps to the look pass
• crossfade - plain mix

step_fps then holds frames to 8-12 fps for a stop-motion / AnimateDiff stutter, independently of the blend (works without a look pass too).

input	type	default	notes
video	IMAGE		The video-model render (H3, LTX...). Its motion and light are kept.
blend	choice (detail transfer, colour + detail, flicker cuts, crossfade, none)	detail transfer	detail transfer: the look pass's fine texture on top. colour + detail: also its palette. flicker cuts: cut to the look pass every few frames. crossfade: plain mix.
amount	FLOAT	0.6	How much of the look pass comes through.
detail_radius	INT	3	Texture scale taken from the look pass.
flicker_every	INT	3	flicker cuts: roughly one cut every N frames.
step_fps	FLOAT	0.0	0 = off. 8-12 gives a hand-made stutter.
fps	FLOAT	24.0	Clip fps, for step_fps.
seed	INT	0	flicker cuts: which frames flicker.
look_pass (optional)	IMAGE		Frames in the look (Feedback Sampler or Restyle). Resampled to the video's length.

Outputs: frames (IMAGE)

Difforum · Symmetry

Difforum_Symmetry

Mirror or kaleidoscope a frame or batch (the Feedback Sampler can also do it inside the loop, where it compounds into a living pattern).

input	type	default	notes
image	IMAGE		An image or a batch of frames.
symmetry	choice (none, mirror_h, mirror_v, mirror_quad, kaleidoscope)	kaleidoscope	Mirror or kaleidoscope the image.
segments	INT	6	Kaleidoscope segments.
mix	FLOAT	1.0	How much of the symmetric image is used.
flip (optional)	BOOLEAN	False	Mirror the other way.
center_x (optional)	FLOAT	0.5	Symmetry centre, 0-1 across.
center_y (optional)	FLOAT	0.5	Symmetry centre, 0-1 down.
angle (optional)	FLOAT	0.0	Rotation of the symmetry axes, degrees.

Outputs: image (IMAGE)

Difforum · Echo Trails

Difforum_EchoTrails

Long-exposure trails across a frame batch.

input	type	default	notes
frames	IMAGE		A batch of frames (a clip).
decay	FLOAT	0.6	How long trails last. 0.9 = long exposure.
mix	FLOAT	0.5	How much of the trail is laid over the frame.

Outputs: frames (IMAGE)

Difforum · Flow Stabilize

Difforum_FlowStabilize

Anti-flicker: blends history along optical flow, only where it matches, so texture stops boiling but motion never ghosts.

input	type	default	notes
frames	IMAGE		A batch of frames (a clip).
strength	FLOAT	0.5	How much history is blended in along motion. 0.3-0.5 calms boiling texture; 0 = off.
flow_scale	FLOAT	0.5	Resolution of the motion estimate. 0.5 = fast, 1.0 = precise.
error_gate	FLOAT	0.15	Where the motion estimate is unsure (occlusions, new content), blending turns off. Lower = safer, higher = stronger smoothing.

Outputs: frames (IMAGE)

Difforum · Detail Guard

Difforum_DetailGuard

Unsharp + contrast + grain, for frames that went soft.

input	type	default	notes
image	IMAGE		An image or a batch of frames.
sharpen	FLOAT	0.3	Unsharp amount.
contrast	FLOAT	1.0	Contrast multiplier.
grain	FLOAT	0.0	Film grain amount.
grain_mode	choice (gaussian, plasma)	gaussian	gaussian = fine film grain, plasma = soft organic noise.

Outputs: image (IMAGE)

Difforum · Upscale (2K / 4K)

Difforum_Upscale

Finish at 2K (or 1080p / 1440p / 4K / xN) for delivery.

With an upscale model (Load Upscale Model: 4x-UltraSharp, RealESRGAN, 4x_foolhardy_Remacri...) each frame is upscaled by the model in chunks, then resized to the exact target; without one it is a clean Lanczos resize. Aspect is kept and sizes stay even for video codecs. Bypass the node (or its group) to deliver at render size.

input	type	default	notes
frames	IMAGE		The clip to deliver.
target	choice (2K (2048 long edge), 1080p (1920 long edge), 1440p (2560 long edge), 4K (3840 long edge), x1.5, x2, ...)	2K (2048 long edge)	Final size. 2K = 2048 on the long edge; xN = multiply.
method	choice (lanczos, bicubic, bilinear, area)	lanczos	Resize filter after the model. lanczos is sharpest.
sharpen	FLOAT	0.15	Light unsharp after resizing.
chunk	INT	16	Frames per model pass; lower it if VRAM runs out.
upscale_model (optional)	UPSCALE_MODEL		Load Upscale Model (RealESRGAN, 4x-UltraSharp, Remacri...). Empty = resize only.
model_use (optional)	choice (auto (when enlarging 2x or more), always, never)	auto (when enlarging 2x or more)	auto: the model runs only when the clip grows 2x or more (640 -> 2K); a clip that is already large (1280 -> 2K) is resized and sharpened, many times faster. always / never force it.

Outputs: frames (IMAGE), info (STRING)