Skip to content

Nodes and their settings

There are five kinds of node. Every flow, ready-made or yours, is built from these.

node palette tab inputs output
Prompt Common - prompt
Image Common - image
Video Common - video
Image generator Generate prompt, image (+ reference, + control) image
Inpaint generator Inpaint prompt, image image
Video generator Generate prompt, image video
Wan video Generate prompt, image video, last frame
Motion transfer (Wan Fun Control) Generate prompt, image, control video
Video with sound (MiniMax H3) Generate prompt, image, reference video, last frame
Output Common media -

Every node's sheet ends with Reset node (all settings back to default, and any batch sweep on it released) and Delete node. Double-tap a node's name in its sheet to rename it.


Prompt

Holds the text. Both boxes are shown on the node itself; tap one to edit it.

setting what it does
Prompt What to draw
Negative What to keep out of the picture (ignored at CFG 1. FLUX.2, Z-Image, Qwen Image)

A token count under each box tells you how much of the model's prompt limit you have used. While you edit a box, a row of buttons under it works on the tag your cursor is in: ↑ and ↓ make it heavier or lighter by 0.1 (written (tag:1.1); back at 1.0 the brackets go), ✕ deletes it, + starts a new tag after it, and the two arrows at the end undo and redo. If the prompt is in Russian or Chinese, a 文A button translates it to English on the phone; the same button undoes it. See Settings → Prompts → Translation. While you type, tag autocomplete offers matching tags under the box (Settings → Prompts → Tag autocomplete).

The picture button on the prompt box writes a prompt from a picture. Pick one already on the canvas (an image node's photo, or what an output node made) or From your phone, and what to write:

writes for model, time
Short one sentence photo models, SDXL, FLUX.2 Florence-2 PromptGen, about 4 s
Detailed a paragraph. Subject, clothes, pose, light, background the same Florence-2 PromptGen, about 4–5 s
Tags Danbooru tags, the character if it knows them, and a rating Illustrious, Anima, Pony WD tagger, about 1 s

It opens on Tags when the prompt feeds an anime checkpoint and on Detailed otherwise. An empty prompt is filled straight away; otherwise you see the result and choose Add after your prompt or Replace your prompt. Each model (548 MB and 378 MB) is offered the first time you need it and listed in Models → Tools. Sentences can get details wrong; for anime, Tags is the accurate one. Nothing is filtered, and nothing leaves the phone.

A prompt with the describe and translate buttons

Image

A photo from your gallery. The empty node offers Choose from photos, which opens your phone's photo picker (on a phone without one, your Gallery app, which may ask for the Photos and videos permission), Choose from files, the file browser for a picture that is not in your gallery, each with its Add button, and Blank page beside them to draw on. It appears on the node straight away, no Run needed. Once there is a picture, a row of square buttons sits above it: Photos (another picture from the picker), Files, Draw (paint over it, see Image to image) and the red bin (clear it), then the file's name.

The Image node before a photo is chosen

Video

A video from the gallery (or Files), for Motion transfer's control: the motion your picture copies. It shows the video's first frame; which part of the video is used, and how it is cropped, is chosen on the Motion transfer node.

Image generator

The node that makes the picture. Its title follows what is wired into it:

wired in title
prompt only Text to image
prompt + image Image to image, or Image edit on FLUX.2 and Qwen Image
prompt + reference Image edit (FLUX.2 and Qwen Image); on SD 1.5 Swap the title stays because the reference is IP-Adapter's picture

The palette card is Image with a chip per family: SD 1.5, SD 1.5 Swap, SDXL, SDXL Swap, Anima, FLUX.2, Z-Image, Qwen Image. The family decides which settings appear.

Scheduler (with its Schedule), Seed and CFG sit under Scheduler, seed and CFG, closed until you open it; closed, it shows what they are set to (Euler A · seed random · CFG 1.5). The seed's padlock above Run still locks the seed without opening it.

Hires fix (or Enlarge ×2), Match colours, Fix faces, Fix hands and Fix strength sit under More options at the end of the node's settings. It opens and closes with a tap and remembers which; closed, it lists the ones that are on.

setting families what it does
Checkpoint all The model. Switching family rewrites the node and keeps its wires. Keeping your steps, CFG and scheduler when asked also keeps the size, pulled inside what the new model can render (SD 1.5 and SDXL take their own sizes)
Steps all Refinement passes. Opens on the model's own recommended value
CFG all How strictly the prompt is followed. FLUX.2, Z-Image and Qwen are made for 1: above 1 the negative prompt is read, but the picture burns quickly
Seed all 0 = a new picture every Run. Type a seed shown on a node to get that one back
Scheduler all The sampling method; a distilled model usually wants its author's choice. SD 1.5 / SDXL: DPM++ 2M, 2M SDE, 3M, 3M SDE, DPM++ SDE (2x slower), Euler, Euler A, LCM. DiT models: Euler (default), Euler A, Heun, DPM++ 2M, DPM++ 2M SDE, ER-SDE, Res Multistep, iPNDM, LCM. Anima: Euler, Euler A
Hires fix SD 1.5, SDXL (not inpaint) After the render, enlarge it ×2 with your first upscaler and redraw the detail with UltraFix: one Run, a picture twice the size. Uses UltraFix's last Steps / Denoise steps. On the Snapdragon 8 Elite: SD 1.5 512 → 1024 about 16 s in all, SDXL 1024 → 2048 about 2½ min. Needs an upscaler: switching it on offers one to download, and so does a Run
Enlarge ×2 every Image node without Hires fix: FLUX.2, Z-Image, Qwen, Krea, Iris, Anima, the Swap models and every inpaint After the render, enlarge it ×2 with your first upscaler, nothing redrawn. The node's own picture is the enlarged one, so wire it into an inpaint (with Only masked) to fix faces at the larger size. E.g. Z-Image text to image → enlarge → an inpaint over the face. Needs an upscaler: switching it on offers one to download, and so does a Run
Fix faces, Fix hands every Image node but inpaint, Iris, Wan and Ideogram (on MOSS-Image since 1.6.222, new: please report) After the render (and Hires fix or Enlarge), each face or hand is found and redrawn zoomed in, with this node's prompt and model, then blended back (ADetailer). A face or hand that already fills most of the picture is left as it is: redrawing it would only soften it. A Swap model redraws it with its LoRAs, without its ControlNet or reference picture. Up to 4 of each, the largest first. Needs the Face and hand detector from Models, Tools: switching it on offers it, and so does a Run. Each patch's tone is matched to the picture along its edge, so it leaves no lighter or darker ring. Each one found adds a render: on SD 1.5 about 2 s, more on bigger models
Detection settings with Fix faces or Fix hands on Opens a sheet with a Faces and a Hands tab, each showing your last picture with what the detector finds: solid boxes are fixed, dashed ones skipped. Confidence (0.30): lower finds small or blurred faces, higher skips things that only look like one. Padding (0.12): how far the redraw reaches past the box. Mask blur: how soft the patch's edge is. Fix strength: how much of a face (or hand) is redrawn, set apart from the other tab; it follows the node's Fix strength until you move it. Saved with the flow
Fix strength with Fix faces or Fix hands on How much of each face or hand is redrawn. 0.4 keeps the pose and fixes the detail
Fix seed with Fix faces or Fix hands on The seed of the fix. 0 = a new face or hand every Run, and a second padlock beside the picture's seed above Run. Lock the picture's seed, leave this one free and Run again: the picture is kept (not rendered again) and only the face is redrawn. Turning Fix faces on, or changing Fix strength, keeps the picture the same way. It cannot carry a face to another picture: the fix redraws the face that is there
Variation seed, Variation strength SD 1.5, SDXL and Anima (not the DiT models yet) A second seed mixed into the starting noise, as in the WebUI: a picture you like, nudged. Strength 0 is the seed alone; 0.05 to 0.3 keeps the picture and changes details; 1 is the variation seed alone. Under More options
Match colours every Image node but Iris, Wan, MOSS and Ideogram Image to image, inpaint and Fix faces / hands come back in the colours of the picture going in, so a redrawn patch has no tinted seam. Does nothing when nothing goes in and nothing is fixed
Schedule SD 1.5, SDXL, the DiT models but Wan and Iris How the noise steps are spaced: Normal, Karras or Exponential (not with LCM); the DiT models also have SGM Uniform, Bong Tangent, Simple, Beta, Beta57, Linear Quadratic and Radiance Shift. Normal is each model's own; on a 4- or 8-step DiT model the other two need about 12 steps
Layout MOSS-Image A tab beside the picture tabs: draw a box for each thing in the picture and say what it is (an object, a character, or text with its exact words). The prompt stays the picture's overall description. MOSS-Image was trained on this kind of prompt, so things land where you put them, and text is spelled as written. A layout written for ComfyUI's MOSS Layout Prompt node can be pasted into the prompt instead. Ported from LinXce's ComfyUI node
Denoise with an image How much of the photo is replaced. 0 keeps it, 1 replaces it. Image to image opens at 0.65, Image edit at 1.0
Resolution SD 1.5 The sizes this model supports. A new size reloads the model on the next Run
Aspect SDXL, Anima; FLUX.2, Z-Image, Qwen, Krea 2, Anima DiT SDXL and Anima: crops the fixed 1024 square to that ratio. The others: sets Width and Height to that ratio. No reload
Width, Height FLUX.2, Z-Image, Qwen 512–2048 in 64-pixel steps. No reload
LoRAs FLUX.2, Z-Image, Anima DiT, SD 1.5 Swap, SDXL Swap Adapters on top of the model, each with a strength. LoRAs
ControlNet tile SD 1.5 Swap, SDXL Swap Type (canny, depth, openpose or none), strength, and the picture. Always an image node wired into control: switching a type on wires the node's own photo in (image to image) or adds an image node (text to image); Replace puts a different picture in its own node, none removes the wire. The node's own photo follows its crop window; any other picture has its own Crop (drawn over the photo, shaped like the render, zoomable out past its edges). An empty picture is skipped with a warning. Canny finds the edges; openpose finds the pose in a photo (the Pose Detector, 19 MB) or uses a skeleton as it is; depth makes a depth map from a photo (the Depth Estimator, 99 MB, about 1 s) or uses a ready-made one. Each ControlNet (SD 1.5 about 370 MB, SDXL about 1.3 GB, 8 Gen 3 and newer) and each estimator is a download offered on the tile and in Models → Tools; a ControlNet not yet built for your chip says so. On an SDXL Swap conversion made without ControlNet the tile is dimmed and says why: of the defaults, Illustrious XL Swap has ControlNet and IP-Adapter (a copy downloaded before 1.6.122 does not. Delete it and download it again) and Juggernaut XL Swap has LoRA only. The tile shows what the ControlNet sees; ⬇ above that picture saves it to the gallery
IP-Adapter tile SD 1.5 Swap (npuforge 1.0.7+ and the defaults), SDXL Swap (npuforge 1.0.12+ with IP-Adapter ticked, and the Illustrious XL Swap default) A reference picture the render follows: an image node wired into reference, added when you pick plus or face (none removes the wire), with its own Crop over the photo (shaped like the render; an empty picture is skipped with a warning). plus carries its subject and style, face a face; strength 0–1.5 (0.6 to start). The encoder (1.3 GB, about 6 s a picture) and each adapter (SDXL's 425 MB) are downloads offered on the tile and in Models → Tools. The tile shows the square the encoder reads. An older Swap model says to convert it again (or, for a default, to download it again)
Crop frame with an image Which part of the photo is used. Tap the picture on the node to frame it
Pad with an image What fills the frame where it runs off a small photo: Black, Blur; edit models also Green
Allow padding FLUX.2, Qwen (edit) Lets the frame zoom out past the photo; the padding is generated
Reference region FLUX.2, Qwen, SD 1.5 Swap, SDXL Swap Which part of the reference picture is used

The bars icon beside Seed, Steps, CFG, Denoise and Scheduler arms that setting for a batch.

Inpaint generator

Everything the image generator has, plus the mask. Families: SD 1.5, SD 1.5 Swap (with its LoRAs and ControlNet; a blend inpaint, so a big fill can leave a soft seam), SDXL, SDXL Swap (with its LoRAs, ControlNet and IP-Adapter; a real inpainting model when converted with Inpaint), Anima, Z-Image, Qwen Image and Krea 2 (the area is redrawn at the Denoise strength and blended back along the mask. Best at 0.5–0.8).

setting what it does
Mask The area to repaint. Inpaint
Draw Paint colours and rough shapes inside the mask; the repaint follows them. It draws on the image node feeding the inpaint, the same layer as that node's own Draw
Only masked On (default): work on a crop around the mask. More detail where you painted
Stitch to original image Off (default): the result is the frame you chose. On: pasted back into the whole photo
Grow Widen or narrow a tapped or picked area
Feather Soften the mask's edge
Enable tap to select Tap an object in the mask editor to select it (Segment Anything, 87 MB)
Enable auto mask Pick Clothes, Face, Hair, Shoes or Bag by name (29 MB); the flow remembers it
Enable Add Objects Paste objects from another photo and repaint only their edges. Add Objects
Pad Black, Blur or Green, for frames that run past the photo (outpainting)

The three Enable boxes remember how you last left them: a new Inpaint node starts the same way.

Video generator

The text- and image-to-video model. Needs the video models (Models → Video).

setting what it does
Seed 0 = a new clip every Run
Upscale On (default): 2× to 1024 × 640. Off: 512 × 320, a little faster
Crop frame With a photo wired in: which part is animated

Wan video

Wan 2.2: text or a photo to a clip of up to 5 seconds. Needs Wan 2.2 (Models → Video).

setting what it does
Seed 0 = a new clip every Run
Seconds 1 to 5, 24 frames a second (Wan 2.2 14B: 1 to 3 at 16 frames a second, 1 by default)
LoRAs Wan 2.2 Turbo (5B) LoRAs, chosen like on the image nodes (Wan 2.2 14B takes none); <lora:name> tags in the prompt work too
Width, Height Any shape in 32 px steps, up to 1280 × 704 worth of pixels (832 × 480 by default; 1280 × 704 takes about three times as long)
Crop With a photo wired in: tap the preview, then drag the frame to choose which part becomes the first frame
Model A dropdown: Wan 2.2 Turbo, Neodragon (the fastest), MiniMax H3 (with sound), Wan 2.2 14B in Image to video only, NSFW Wan 1.3B in Text to video only; the node becomes that model's node and back. Swap flow on the node goes between Text to video and Image to video with the model kept (how)

Its last frame output is the clip's final frame as a picture: wire it into another Wan video node's image to continue the clip with a new prompt. That node's video is then the whole film so far, both clips joined (longer clips).

Video with sound

MiniMax H3: a prompt into a clip with its sound (speech, sound effects, music), up to 5.2 seconds. Needs MiniMax H3 (Models → Video). Its prompt is written in H3's sections, who speaks, the scene and the line, the sounds, the music (how). A picture on image is the clip's first frame; a picture on reference is a subject the clip keeps in a new scene (MiniMax H3 Reference, its own download, 1.5 to 2.5 times slower), one or the other.

setting what it does
Width, Height 384 to 1344 in 32 px steps, up to 768 × 1344 worth of pixels (832 × 480 by default; 384 × 384 is the quickest)
Seconds 0.9, 1.6, 2.3 (default), 3.0, 3.8, 4.5 or 5.2, 24 frames a second; a larger size fits fewer
Smooth motion Twice the frames, as on the Wan nodes
Seed 0 = a new clip every Run

Powered by MiniMax H3 (MiniMax H3 Community License: not for the EU, the UK, South Korea or the US).

Motion transfer

Wan 2.2 Fun Control: the picture wired into image moving like the video wired into control (a Video node). Needs Wan 2.2 Fun Control (Models → Video).

setting what it does
Seed 0 = a new clip every Run
Width, Height As on Wan video
Mode Pose or As is: how the video becomes the motion. The ⓘ says which to use when (Motion transfer)
Crop Which part of your picture is used, as on Wan video (with Line up my picture on, the app crops it to match the dancer instead)
Frame rate 24, 12 or 8: how often the video is read. Lower is faster, with bigger jumps between frames
Follow the person The crop moves with the person on every frame (on by default)
Line up my picture Your picture's person starts where and as large as the dancer (on by default)
Keep audio The video's sound over the same part goes into the clip (on by default)
Smooth motion Twice the frames, as on Wan video
Video Tap the preview. Two handles choose the part of the video that drives the clip; its length is the clip's (up to 30 s, made in pieces past one pass). With Follow the person off, a box on that part's first frame, at the clip's shape, cuts every frame the same way

It has no Seconds (the part of the video sets the length) and takes no LoRAs.

Output

Where the result lands. It shows the picture or clip; tap it to see it full screen. Its finishing steps sit under Post-processing at the end of its settings, folded until you open it (closed, its header names the ones switched on). It remembers whether you left it open, apart from a sampler's More options. There are three cards, Upscale, Colour correction and Adjust colours, each with:

  • After every Run: a switch that applies the step to every new picture.
  • A button. Correct applies colour correction to the last picture right now. Upscale and Adjust open a window with that step's settings (the upscaler and scale, or the four sliders) and a button that applies it to the last picture. Nothing is rendered again and no seed is locked: the flow reruns on that picture's own seeds, so only the Output works.
  • Under the name, what Upscale and Adjust colours are set to (for example Anime 4x · 4x), so the settings in the window are never out of sight.
setting what it does
Autosave On (default): every Run is kept in Results with the flow that made it
Upscale Enlarge the result before keeping it. The node then shows what it received above what it made. Greyed out on an output wired a video: a clip cannot be upscaled
Colour correction Fixes a dull or tinted picture: darks darker, brights brighter, a colour cast removed, after any upscale. A warm scene stays warm
Adjust colours Brightness, Contrast, Saturation and Warmth sliders (each -1 to 1, 0 unchanged) in its window, after Colour correction
Upscaler In the Upscale window: which upscaler to use. Install them under Models → Upscalers
Scale In the Upscale window: 2×, 3× or 4×. A result that would pass 4096 px gets the largest scale that fits
Save (video) Also write the MP4 to Movies/Nightmare