Documentation

MiniMaxDirector Upscale — the second stage. The Upscale switch in the Upscale group renders the clip as usual, then enlarges the latent and refines it. Off, nothing below it runs and the graph is exactly the one-pass director render. The order matters: render first, look at it, and only then pay for the resolution. Flip Upscale on and queue the same graph again — nothing upstream changed, so ComfyUI serves stage 1 from its cache and only the upscale and the refine pass cost anything. That cache lives in the running ComfyUI, so a restarted pod re-renders stage 1 (identically — same seed, same graph).The node's mode is megapixels, and megapixels is the size you land on, not a multiplier — the canvas defaults to 1344×768, which is 1.03 MP, so 2.0 is a bit under 1.5× per side. It must be larger than the canvas; the node refuses to shrink. align 32 is a pixel grid, and 32px is one 2-latent step: it keeps both latent axes even, which is what the DiT's 2×2 patching wants. keep_proportion off for the same reason — with it on only the width lands on the grid and the height follows the aspect ratio, which can round to an odd latent row. H3's latent holds picture and sound together, and the upscaler takes a plain video latent — handed the joint one it raises 'NestedTensor' object has no attribute 'dim'. So the branch splits the audio off (Separate AV Latent), upscales the video, and joins them again (Concat AV Latent) before the refine pass. Both are core ComfyUI nodes, and it is the same wiring the pack's own example workflows use. The refine pass is an 8-step schedule split at step 4, so it starts at sigma 0.9231 and takes 4 steps down. That is a real second render at the larger size — **it costs time and VRAM, not just the 691MB upscaler**. Fewer refine steps is cheaper and softer; more is slower and sharper. The upscaler is LBH-123-AI's H3 latent upscaler, working directly on H3's 24-channel latents so nothing round-trips through the 5B VAE. Days old at the time of writing, and unproven at our sizes — check the result before trusting it with a long clip. The pack rewrote itself on 2026-08-19: H3LatentUpscalerNodeMegapixels was deleted and MinimaxH3LatentUpscaler3D took its place, with align moving from latent units to pixels. The onstart pins the clone to a commit for that reason — an unpinned one broke this graph on a pod mid-day. This one branch is the graph's only third-party dependency, and this pod already has it. Off a pod it is git clone https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler.git in custom_nodes, plus minimax_h3_latent_upscaler_3d_fp16.safetensors in models/latent_upscale_models/. Without it ComfyUI opens this graph with a Missing Node Types dialog and one red node — expected, and harmless while Upscale is off. --- Speed — one toggle. The Turbo switch in the Speed group picks the render mode; nothing else in the graph changes.The turbo LoRA is lightx2v's Ref2VA 4-step distillation (docs: ModelTC/Minimax-H3-Turbo). It is distilled for 4 steps at 544p on the shifts H3 already defaults to (video 12 / audio 3), so leave strength at 1.0, the scheduler on simple, and the step count where the toggle puts it. Ref2VA turbo is a v0.1 preview — audio and fast motion are its weak spots, and the distillation was trained against the bf16 base while this bundle runs the pruned int8 one. Switch Turbo off for a final render. --- Lay out shots on the Director node's timeline; it compiles them into the single structured prompt MiniMax H3 reads, and keeps the clip on a length H3 accepts (length % 17 == 5 at 24 fps). Why that rule: the model denoises a latent whose time axis is a row of slots, and the video VAE packs 17 frames into 5 of them (after a 5-frame head worth 2). A length off the lattice would need a fraction of a slot. So 124 f is legal, 130 f is not, and only 8s, 25s and 42s land on whole seconds. • Two panels beside the director show what was built and what the linter thinks, both updating as you type rather than after a run -- a warning that arrives after the render arrives after the cost. • Drop an image on a shot with Add Image; it becomes <Picture 1> automatically. • Models: the H3 bundle (ref2va unet, Qwen3-VL text encoder, video + audio VAEs). • The title bar carries the pack version and build date at its right end. The three prompt buttonsAdd Image / Add Audio / Add Video attach a real file instead, and the prose is given a <Picture n> / <Audio n> / <Video n> token pointing at it. With a block selected that carries no file yet, the file lands on that block; otherwise it gets a block of its own. An attached audio or video takes the span it actually runs for -- the file is measured before its block is placed -- bounded by the block after it and by the end of the clip. Files, under the transport row beneath the tracks, opens the list of every file the clip carries -- the ones on blocks, with the token they compile to, and the ones on no block at all, shown dashed. Its + file adds a file the clip carries with no moment of its own -- any of the three kinds, taken from the file. A block says "this stretch of the video is about this file" and compiles as (appears in [Shot n]); a face to be carried onto whoever is on screen is about no stretch, and putting it on a block cuts the clip at a seam the model then acts on. An unplaced file is numbered with the rest, described on a card, and written into any prompt by its chip. Drag it out of the list onto a track and it becomes an ordinary block at the frame you dropped it; x takes it off the clip. Every chip drags, placed or not: an unplaced file moves onto the track, one already on a block is copied, which is how the same photograph is used in two shots without going back to disk for it. Copying, and the keyboard Clear empties the piece: every block, the global prompt, the music and every card on WHO & WHAT. One undo step puts the timeline back; the cards do not come with it.The playhead The red line is where every Add button puts its block -- click the empty part of a track to move it, or drag the scrubber. If it is standing inside a block there is no room, so the new one goes on the end instead. • Dragging a block or its edge snaps to the playhead and to the edge of every other block, on any track, within a few pixels -- so a cue can start exactly where a shot does. • S cuts the selected blocks in two at the playhead. The second half keeps the prose and drops any attached file, so the same picture is never in the prompt twice. • Zooming with + / - recentres the view on it. The clip settings, left to rightA block that grows the clip -- added, dragged past the end, or given a longer length -- takes the lattice padding itself, so the timeline is exactly what will be generated and no frame of the output is left without a shot describing it. The editor opens at fit, with the whole clip on screen. The tabs The panel under the toolbar shows one of four things, and remembers which one across a reload, along with the block that was selected: TIMELINE (the tracks and the selected block's fields), WHO & WHAT (one card per thing the prompt names, with the count on the tab) and GLOBAL (the two clip-wide prompt boxes). Those three are where the piece is written, so they sit together on the left. IMPORT / EXPORT, at the right-hand end of the row, is what you do with the piece once it is written. The node is exactly as tall as whatever panel is open -- nothing here is a fixed height that clips. The card list is the one exception: drag the grip in its bottom-right corner, or the node's own corner while WHO & WHAT is open, and the height you set is stored on the node and comes back with the workflow. Import and export IMPORT / EXPORT holds the whole piece as one JSON -- the timeline, the cards on WHO & WHAT and the clip's own width, height and default resize -- with four buttons: Save file opens the browser's own save dialog -- the folder and the name are yours, and it suggests minimax-director-<date>.json -- Load file reads one back, Copy puts the same JSON on the clipboard, and Paste opens a box -- press Cmd/Ctrl+V in it and the piece loads as it lands, which is one action and no browser permission popup. Firefox and Safari have no save dialog and download the file instead; the line beside the buttons names what was written either way. A load replaces the node -- timeline, cards and settings -- and asks first when there is anything to lose. Cmd/Ctrl+Z puts the timeline back; the cards are a document of their own and the undo stack does not hold them, exactly as Clear says. The JSON names the files, it does not carry them: a picture's filename, never its pixels. So a load ends by asking this ComfyUI which of the named files it actually has, and lists the ones it does not with an Upload button -- pick them from disk and every block pointing at each name is re-pointed at the uploaded copy. The same check runs whenever the tab is opened, so a workflow somebody sent you says what it is missing without being imported at all. A file this ComfyUI does not have is drawn in red wherever it appears -- the block, the chip under the prompt, its row in Files, and the card's face and from -- with the count on the tab itself, IMPORT / EXPORT · 1 missing. Red means broken here; amber still means unfinished. While one is missing the panels are locked, and clicking one flashes the blocks whose file is gone. Three things stay live because they are the ways out: re-upload -- in the middle of the block itself, where the picture would be, and on the file's row in Files -- and Delete and Clear for when the answer is that the block should go. Any file off disk will do for a re-upload -- renamed on disk is the usual reason one goes missing. The run is refused as well, before a frame is sampled, naming the files and what to do about them. That is the check that holds: the lock is a browser drawing a warning, and a queue from another tab or from the API never sees it. The segment panel, under the timeline Select a block first -- with nothing selected there is nothing to edit and the fields are not on screen. Everything here edits that block.Select several blocks and the panel becomes a selection panel: only the fields that apply to all of them (motion / strength / speed, enter with, used as, keep file), each starting on leave as is. same length and close the gaps are always there -- a frame count means the same thing on every track. For shots there are two more: merge into one shot, which is what MiniMax asks for when a cut only changes the distance, and make the speech continuous, which writes one sentence across the cuts. Frames first, seconds after, everywhere: the playhead clock reads 48 f = 2.00s and the selected block Start: 0 f | End: 96 f | Length: 96 f = 4.00s. Frames are what the document stores and what H3 is given; seconds are the translation. Every number box -- start, end, length, duration, width, height, same length -- takes effect on Enter or when you leave it, not as you type, so one can be cleared and retyped without the half-finished number being read and refused; what lands in the box afterwards is what was actually set. Enter finishes any field and leaves it, prompt boxes included, and leaving a box flattens what is in it: paste a paragraph and it collapses to one line, because one line is what the compiled prompt carries. A dialogue row with nothing typed in it dims -- the row and its background both -- because the compiler ignores it until it has words. Along a block's bottom edge sit its chips: the file it carries (IMAGE · face.jpg), and one per transfer taken out of that file, FACE -> SPEAKER, amber FACE -> ? while nobody has been named to receive it. The GLOBAL tab holds the two fields set once for the whole piece:Paste freely. Line breaks are structure in the compiled prompt -- subject_definitions and retention_analysis list one entry per line, and a blank line starts a new field -- so every box is flattened to a single line on the way out. A paragraph pasted from a document arrives as one sentence, not as a subject nobody wrote. Why seconds are read-only. Every greyed = ... s box is a reading, not an input. A second is 24 frames wide, so a block typed as 1.08 came back as 26 frames and was shown as 1.08 again -- the number actually set was never on screen. The clip is written in seconds and cut in frames, and only one of those can be the field you edit. Dialogue H3 makes the voice and the picture in one pass, and the guide's form for it is exact: The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d> The editor writes it for you, and splits it the way the work actually splits: **who the people are is written once in the WHO & WHAT tab, and a block only says who talks and what they say**. WHO & WHAT is one card per thing the prompt has to name -- usually a person, but equally a costume, a prop, a place or a style, which fill in the same card with the voice row left empty. Several cards may point at one file: that is how a single photograph names several things, each numbered separately. S1…Sn is a speaker -- who says a line; any card with a voice. <Subject 1…n> is a subject -- a person, a costume, a prop, a place, a look the model must keep. A card is one only with a file and a description; without a file there is nothing for the prompt to point at, so it can only be a voice. Both tokens are MiniMax's, and a card can be one, the other, or both. The block's FILE row lists the subjects drawn from that file, one per line, with edit beside each and + another card underneath -- which is how a single photograph names a person, their coat and the room behind them. The subject chips write the token for you. They sit along the bottom of the prompt box itself -- every numbered card a chip, its file's thumbnail beside <Subject 2> suit; clicking one splices the token into the box above it where the caret is, and a chip that box already names is lit. GLOBAL PROMPT carries the same strip. Typing the number by hand is the alternative, and getting it wrong is silent -- the prompt cites a subject that does not exist and nothing on screen says so. Name a subject in every shot it appears in. That is how one basket stays one basket across a cut, rather than three descriptions of a basket: the same <Subject 1> written into all three shots, by chip. The compiler follows -- its line in retention_analysis reads (appears in [Shot 1], [Shot 2] and [Shot 3]) instead of naming only the shot its file sits on, and without it the model is told the basket belongs to one shot and is free to invent another for the next. The files have chips too, dashed, after the subjects: <Picture 2> face.jpg, <Audio 1> voice.mp3, <Video 1> clip.mp4, one per file on the timeline. The compiler writes a file's token into its own block's line; pointing at it from anywhere else -- a recording the mouth has to follow, a picture a later shot refers back to -- is what these are for.A card that is doing nothing looks like it. A card counts when it names a file and says what that file is -- that is a <Subject n> -- or when it describes a voice something actually speaks. Short of either, the compiled prompt is byte-for-byte what it would be with no card there, so the card goes flat: transparent, dashed, dimmed, with the reason in amber across it and the same line in report.Two badges say the rest. <Subject n> is what the prompt will call this card, and a hollow no <Subject> where it would be means no file was picked. A green [Shot n] says where the card is heard, which is otherwise only visible from the TIMELINE tab. Add adds a card; they speak switches dialogue off for the whole clip -- every row and every <d> at once, cards kept. The voice row goes with them, `voice from included, and so does the Sn` badge: with nobody speaking a timbre reference instructs nothing and the compiler drops it, and no card is called by a speaker number. Describing the same speaker two different ways in two shots used to be possible; to the model that reads as two people wearing one label. On the block:+ line adds another row, so one block can hold a conversation: a line each, spoken in turn, compiled as one <d> apiece. It goes dead -- dashed and dimmed, with the reason in amber -- while a row on the block still has no words, since the compiler ignores that row and a second empty one adds a second nothing. The red bin at the end of a row removes it -- the same delete button the subject cards carry. Clicking a face hands the line to that person alone; hold Cmd or Ctrl to add another, and another -- the row says so beside the faces, and any number of them can say the words at once. Three readings, kept apart on purpose: a chorus is one row with two faces, a conversation is two rows, and an argument -- overlapping speech with no agreed words -- is neither. Write that one in the segment prompt and put the sound in an AUDIO cue; there is nothing for H3 to quote. Attached files: used as, describes and keep file used as -- what the file is for. The picker offers only what the file could sensibly be: a picture gets reference, storyboard and the three frame anchors; a video gets reference, continue from and edit; a recording gets reference alone, and with nothing to choose the control is not drawn. A document already holding some other combination keeps it, shown until you change it. It decides the task type the summary opens with, and the guide wants every relationship named:Only first frame and last frame have an input on the model. keyframe is the same idea one step weaker: MiniMax's guide counts it as a frame anchor, but the core node takes exactly two stills -- first_frame and last_frame -- so a picture that should be a frame in the middle has nothing to be plugged into. That block's image travels with the references and the prompt asks for the placement in words -- <Picture 1> ([Shot 2] keyframe) where a reference would read (appears in [Shot 2]). An end is a guarantee, the middle is a request the model follows loosely. A keyframe is fitted to the clip, not the other way round. A block used as first frame or last frame carries a fit picker beside keep file. crop, the default, scales the picture and cover-crops it from the centre: proportions survive, an edge is lost. stretch hands it over untouched, which is what ComfyUI does on its own -- every pixel kept, the picture squashed. A picture already of the clip's shape is untouched either way; when the shapes disagree the report names both sizes and what it cost. To keep the whole picture, give the clip the picture's shape, or attach the file as a reference -- that path scales without cropping and lets the model compose the rest of the frame around it. The settings row's default resize sizes reference pictures only -- a reference video is sized by its own rule, a keyframe by fit. It stays live whatever the clip carries. A picture in the Files list counts before it is placed: it reaches the model in the same reference list as one on a block. A picture answers for itself with its own resize, offered both on its row in Files and on a block's FILE row -- it belongs to the file, so both write the same thing. The clip's value is only for the pictures that say nothing: max on the face you have to keep, match on the mood board behind it. What it trades is detail against time. A reference picture becomes tokens the model reads beside the prompt, and those tokens are re-read at every sampling step -- more pixels, finer detail, more time. match shrinks it to about the clip's pixel count: fast, enough for a scene, a style, a mood. max allows 2048 px on the short side: slower, and what keeps a face the same face. Neither enlarges a picture or changes its proportions. default resize does not touch width and height. The clip used to take the shape of the first reference picture whenever it said match; now a picture block carries set width & height beside detach media, which does it on request -- for a keyframe too, which is what the crop warning asks for. An attached audio adds audio reuse or audio reference depending on its keep. Several at once combine: [keyframe completion + video continuation + audio reuse]. A segment holding a real file switches the prompt into H3's full-reference format -- six sections instead of three -- and gets its own row of fields. describes -- read-only, and there is no box here. What a file is is written once, on a subject card, and this line shows that card's sentence beside the <Subject n> it became with a link to the WHO & WHAT tab. Until you add one it reads `nothing describes this file yet`, and the linter says the same: an unnamed reference is one H3 has to guess at. Why not a box on the block? Because a file used to define something is cited inside that thing's definition rather than given a line of its own -- MiniMax's own rule -- so a second box here would have been a field the prompt threw away, which is exactly how it behaved. One file, one description. A frame anchor carries no description box. The picture goes to the vision encoder with the prompt (clip.tokenize(prompt, images=...) in core's nodes_minimax_h3.py), so saying what is in it tells the model nothing it cannot see -- and every source for such a sentence is wrong somewhere: a filename says nothing, and the shot's own prose is the motion across the shot, which a last frame does not contain. An anchor names the frame it is, in subject_definitions and in retention_analysis alike. Cards still work on one. keep file -- how much of the file survives into the video. One per file, always; it also sits on the block itself, bottom right. A subject card drawn from this file carries its own keep it for the thing, which may differ. The sentence it produces lands in retention_analysis, as <Picture 1> (appears in [Shot 2]): fully_preserved - the raccoon, ...These are H3's own words, not ours. It reads them as instructions, so a wrong one is worse than a vague describes: fully_preserved on a style reference asks the model to reproduce the whole frame. What a reference video actually carries. Motion, reliably. Grain, grade and fine texture, not -- and where a clip's grade does land it overrides the light your words asked for, so a sunny scene in the prompt beside an evening clip on the reference list is a fight your words tend to lose. Ask for a film stock in words and keep the video for movement. The same trap in miniature: fully_preserved on a photograph taken at dusk brings the dusk along with the animal; partially_preserved keeps the animal and lets the light change. A video's own soundtrack is a passenger. It is wired in beside the frames and numbered whether or not you wanted it, so the compiler leaves it out of the prompt entirely unless something you wrote names its token -- a clip attached for its motion is not a clip whose sound you asked for. Named, it is described as the audio in that file, never in the words you wrote about the picture. A clip used as a continuation or a frame anchor keeps its entry: carrying the sound over is the point there. An audio file is graded in its own words, because H3's format defines a different set for sound: fully_copy (reproduce this recording), partially_copy, reference (only the timbre or texture is followed), weak_reference. The picker follows the file, so there is nothing to get wrong. No marker copies the file's samples into the clip. A reference audio is encoded into the conditioning, and the soundtrack that comes back is the one the sampler produced and VAEDecodeAudio decoded -- fully_copy asks H3 to re-perform the recording, and how close it lands is the model's business. To ship the recording itself, wire it into CreateVideo in place of the decoded audio. A card that reaches the prompt as nothing is called out too: a card names a file, which makes it a <Subject n>, or describes a voice, which becomes the words in front of (S1). With neither, the compiled prompt is byte-for-byte what it would be with no card there, and the row on screen says so. The reverse is called out too: a voice nobody speaks with, where no line names that card's S -- an instruction about how somebody sounds, applied to nothing. Subjects live on subject cards. Point a card's from at a file and it becomes a <Subject n> of its own, tracked apart from the picture it came from: <Subject 1> is the man's face, from <Picture 2>. That separation is what a face swap needs. The picture stays a weak_reference -- you do not want the whole frame back -- while the card's keep it is attribute_transfer onto the person in another shot. The shot then mentions <Subject 1> rather than <Picture 2>, because naming both asks for two different things at once. onto on the card names who receives the face, and the block carrying the picture shows the move as a chip. A picture whose only job is defining somebody gets no <Picture n> entry of its own: MiniMax's guide asks for it cited inside the <Subject n> line instead. An image used as a first frame or keyframe keeps its entry either way -- it is a real frame of the video, whoever else it defines. The block's FILE row shows that definition read-only, labelled with the <Subject n> it became and linked to the card. When the role does keep the file an entry -- a frame anchor, an edit source -- the card's sentence fills that in as well. Camera moves A move is three choices, the way MiniMax documents it: motion type, amplitude, speed. H3 reads prose, not enum values, so the three become one sentence.A zoom and a push-in are not the same move: a zoom changes the focal length with the camera standing still, a push-in moves the camera body. The model knows the difference. amplitude (small / large) and speed (slow / fast) are added when set -- *The camera pushes in with small amplitude at slow speed.* Both default to medium and normal, which the guide writes by leaving them out, so those options add nothing on purpose. A note typed into a camera segment is appended after the sentence, so write it as a continuation rather than a sentence of its own. Camera work is its own block because a move can straddle a cut -- merging it into the shot line would silently pick a side. What the linter now checks report warns, never refuses: a description outside the 350-500 words MiniMax asks for on a generation task; two adjacent shots that describe the same thing at a different framing (the guide asks for a camera move, not a cut); an empty AUDIO track, because overall_soundscape: N/A tells H3 the clip is completely silent; a voice reference asked to be copied; more reference material than H3 takes -- 9 images, 3 videos, 3 audio, 15s of video and of audio, 12 files in all; a picture outside 256-5760px or outside a 0.4-2.5 ratio -- that one and a clip outside 2-15s are refused when the file is picked, before anything is placed, and the three media buttons go dim once their bucket is full; a line marked carries over with nothing after it; and a guessed word where the guide wants [unclear].

MiniMax H3 MiniMax H3 is MiniMax's general-purpose, omni-modal generation model. It jointly understands text, image, video, and audio, and generates video with native stereo audio: voice, sound effects, and music are modeled jointly in a single forward pass, not layered on afterward. Output is up to 2K resolution, 24fps, and up to about 15 seconds. About this workflow This template runs the Image to Video task (MiniMaxH3ImageToVideo node), which covers both: • t2va (text-to-video), when no images are connected • fl2va (first/last-frame image-to-video), when first_frame and/or last_frame are connected Key inputs • first_frame / last_frame: optional keyframes; the model generates the motion between them • prompt: describe the shots, motion, and the accompanying audio (dialogue, SFX, music) in one block • width / height: set via Resolution Selector. H3's native canvas is a 768px short edge, capped at 768x1344 pixels, rounded to a multiple of 32 • duration (seconds): converted to a valid frame length by the Math Expression node, snapping up to the model's 17-frame-per-block (17k+5) grid at 24fps

MiniMax H3 MiniMax H3 is MiniMax's general-purpose, omni-modal generation model. It jointly understands text, image, video, and audio, and generates video with native stereo audio: voice, sound effects, and music are modeled jointly in a single forward pass, not layered on afterward. Output is up to 2K resolution, 24fps, and up to about 15 seconds. ComfyUI links • ComfyUI#15224 • 🤗 Comfy-Org/MiniMax-H3 About this workflow This template runs the reference-to-video (ref2va) task using the MiniMaxH3ReferenceToVideo node. It takes any mix of reference images, videos, and standalone audio, and weaves them into the generation to lock in a character's identity, a style, a motion, a camera move, or a voice. Key inputs • ref_images / ref_videos / ref_video_audios / ref_audios: up to 9 reference images, 3 reference videos (each may carry its own paired soundtrack), and 3 standalone reference audio clips • prompt: reference the inputs by tag, in the exact order they were connected, for example <Picture 1>, <Video 1>, <Audio 1>, then describe the target scene, motion, and audio • ref_image_size: match scales references down to the generation's resolution (faster); max keeps up to a 2048px short edge for stronger identity fidelity, at the cost of speed since reference tokens ride along every sampling step • width / height: set via Resolution Selector. • duration (seconds): converted to a valid frame length by the Math Expression node Sampling and decode • Sampler: res_multistep. beta or normal scheduler tends to outperform simple for reference-heavy prompts like this one • The sampler's joint audio+video LATENT output feeds directly into both VAEDecode (video, minimax_h3_video_vae_fp16) and VAEDecodeAudio (audio, minimax_h3_audio_vae_fp32); each decode node automatically pulls its own half out of the packed latent. CreateVideo then muxes the two into a single MP4 with synced sound • The diffusion model here is minimax_h3_ref2va_pruned_int8_convrot.safetensors, a different set of weights from the fl2va model used by the t2v/i2v templates Ref2va's output is very sensitive to prompt wording; matching the reference tags precisely and being explicit about which reference drives which part of the shot tends to work best.

MiniMax H3 MiniMax H3 is MiniMax's general-purpose, omni-modal generation model. It jointly understands text, image, video, and audio, and generates video with native stereo audio: voice, sound effects, and music are modeled jointly in a single forward pass, not layered on afterward. Output is up to 2K resolution, 24fps, and up to about 15 seconds. ComfyUI links • ComfyUI#15224 • 🤗 Comfy-Org/MiniMax-H3 About this workflow Key inputs • prompt: describe the shots, camera moves, and the accompanying audio (dialogue, SFX, music) in one block • width / height: set via Resolution Selector. H3's native canvas is a 768px short edge, capped at 768x1344 pixels, rounded to a multiple of 32 • duration (seconds): converted to a valid frame length by the Math Expression node, snapping up to the model's 17-frame-per-block (17k+5) grid at 24fps

How to use Lip-sync an existing clip to speech you supply. 1. Load Video (red) — upload the clip whose mouth should move. 2. Positive Prompt (red) — describe the shot and write what the person says. 3. IC-LoRA (orange) — ltx-2.3-22b-ic-lora-dubit-0.9, strength 1.0 is the tested value. 4. Press Run — stage 1 generates at low resolution, stage 2 refines it. The audio comes from the loaded video. To dub a different voice, mux your voiceover onto the clip before uploading it.

SCAIL-2 — Character Animation Take the motion out of one video and put your own character into it. The driving video's background is not kept — the scene is generated fresh around your character. Fill these in 1. Load Video — your driving clip. Only the movement is used, never the appearance. 2. Load Image — the character to animate (person, mascot, drawing). One character only. The output video is sized to this image, so a portrait photo gives a portrait video. 3. Run SAM3 Video Track — upper node, fed by Load Video. Its text box names the subject to copy motion from. One person in the clip: leave human. Several people: pick one, e.g. man, woman in a red dress — otherwise SCAIL-2 gets two motion tracks and one character, and the result breaks. 4. Run SAM3 Video Track — lower node, fed by Load Image. Leave at human for a person. For a non-human character use its noun, e.g. dog, robot. 5. CLIP Text Encode (Positive Prompt) — describe your character and the action, e.g. a short-haired man in a striped shirt, hands on his hips, full body. Add full body if you want legs in frame. 6. Press Run. Check the masks first The two Preview Image nodes show the tracking masks. Exactly one subject should be coloured in each. If extra subjects light up, make the prompt in step 3 or 4 more specific, or raise detection_thres above 0.50. Do this before any long render — a bad mask wastes the whole run. Length Default is 81 frames at 16 fps — about 5 seconds, taken from the start of your clip. To go longer, raise these two together and keep them equal: • Load Video → frame_load_cap • Wan SCAIL To Video → length 161 ≈ 10 s, 321 ≈ 20 s. SCAIL-2 is trained at 81 frames, so longer runs cost more VRAM and the character may drift. To start somewhere other than the beginning, set Load Video → skip_first_frames (in frames, at 16 fps — 160 skips 10 s). Sound This workflow does not do sound. The render is always silent — SCAIL-2 generates picture only, and nothing in the graph carries audio through to the output. Add the soundtrack afterwards in a video editor, using your original clip as the audio source. Trying to attach it here is not worth it: the render is a short slice of your clip, so the audio would not line up anyway. Video format The upload button accepts .mp4, .webm, .mkv and .gif. H.264 MP4 is the safe choice. If a clip is rejected with "Invalid video file", re-encode it before uploading. A common cause is a movie-rip audio track (AC-3) that the pod's ffmpeg cannot decode: ``` ffmpeg -i input.mp4 -c:v libx264 -pix_fmt yuv420p -an clean.mp4 ``` -an drops the audio, which this workflow does not use anyway. Keep the clip's own resolution — it is resized internally, so a huge 4K source only costs upload time. Leave alone unless you know why • Negative Prompt — a fixed quality filter, not something to describe your video with. • KSampler — steps 6, cfg 1.0, euler / simple. These are tuned for the distilled LoRA; raising steps or cfg makes it worse, not better. • Create SCAIL-2 Colored Mask → replacement_mode — false here on purpose. Setting it true switches to the replacement behaviour (keeps the original background), which is what the scail2-replacement workflow already does. • Every model this workflow needs is already installed on the pod.

SCAIL-2 — Character Animation (two characters) Same graph as the single-character animation, driven by a clip with two moving subjects. Only the inputs and the prompts differ. There is only ONE Load Image — and that is correct There is no second image node, and you should not add one. Both characters come from a single reference image that already contains both of them. SAM3 finds both figures inside that one picture and hands SCAIL-2 two separate coloured regions. It has to be one real photograph of both subjects together — one background, one camera, one light. The output video is built from this frame, so whatever you hand over becomes the scene. Do not glue two separate photos side by side. A collage keeps both backgrounds and the seam between them, and the render comes out looking like two videos in one frame. If all you have is a separate photo of each character, this workflow cannot merge them — run the single-character scail2-animation workflow on each one instead. Fill these in 1. Load Image — one picture holding both characters. The output video is sized to this image, so a wide image gives a wide video. Both characters should be clearly separated and, if you want limbs animated, fully in frame. 2. Load Video — a driving clip with two moving subjects. Their motion is copied; their appearance is not. 3. Run SAM3 Video Track — upper node, fed by Load Video. Leave at human when the clip has exactly two people and you want both. Use a narrower word only if there are extra people to exclude. 4. Run SAM3 Video Track — lower node, fed by Load Image. Leave at human for two people. For non-human characters use a word that matches both, e.g. mascot or character — a word matching only one of them will leave the other untracked. 5. CLIP Text Encode (Positive Prompt) — describe both characters and what they do together, e.g. a black dog mascot character and a green-and-cream bird mascot character holding hands and dancing on a white stage. 6. Press Run. Check the masks first — this matters most here The two Preview Image nodes show the tracking masks. You need two differently coloured regions in each: two in the driving mask, two in the reference mask. If either shows one region, or three, fix the prompt in step 3 or 4 before rendering. Who maps to whom is decided by Create SCAIL-2 Colored Mask → sort_by (area by default — biggest region first in both masks). If the wrong character gets the wrong motion, that pairing is the reason. Length Default is 81 frames at 16 fps — about 5 seconds, from the start of the clip. To go longer, raise these together and keep them equal: • Load Video → frame_load_cap • Wan SCAIL To Video → length 161 ≈ 10 s, 321 ≈ 20 s. SCAIL-2 is trained at 81 frames — longer runs cost more VRAM and drift more, and two characters drift faster than one. To start later in the clip, use Load Video → skip_first_frames (frames at 16 fps — 160 skips 10 s). Sound This workflow does not do sound. The render is always silent — SCAIL-2 generates picture only, and nothing in the graph carries audio through to the output. Add the soundtrack afterwards in a video editor, using your original clip as the audio source. Trying to attach it here is not worth it: the render is a short slice of your clip, so the audio would not line up anyway. Video format The upload button accepts .mp4, .webm, .mkv and .gif. H.264 MP4 is the safe choice. If a clip is rejected with "Invalid video file", re-encode it before uploading. A common cause is a movie-rip audio track (AC-3) that the pod's ffmpeg cannot decode: ``` ffmpeg -i input.mp4 -c:v libx264 -pix_fmt yuv420p -an clean.mp4 ``` -an drops the audio, which this workflow does not use anyway. Keep the clip's own resolution — it is resized internally, so a huge 4K source only costs upload time. Leave alone unless you know why • Negative Prompt — a fixed quality filter, not a place to describe your video. • KSampler — steps 6, cfg 1.0, euler / simple, tuned for the distilled LoRA. Raising steps or cfg makes it worse. • replacement_mode — false here on purpose; the background is meant to be generated fresh. To keep an original background instead, use the scail2-replacement workflow. • Every model this workflow needs is already installed on the pod.

SCAIL-2 — Character Replacement Swap one person in your video for your own character. The original scene, background and everyone else stay exactly as they are. Steps 1. Load Video — upload your clip. The output is sized to this video, not to your image. 2. Load Image — upload the character who takes their place. A clear, full-body photo works best. 3. Run SAM3 Video Track (the upper one, fed by Load Video) — its text box says who gets replaced. One person in the clip: leave human. Several people: name the one you want, e.g. man or woman in a red dress. 4. Run SAM3 Video Track (the lower one, fed by Load Image) — leave it at human. 5. Positive Prompt — describe your new character inside the video's setting, e.g. bearded man in a grey suit sitting at the desk. 6. Press Run. Check the masks before a long render The two Preview Image nodes show the tracking masks. In the driving mask, only the person being replaced should be coloured. If extra people light up, make the prompt in step 3 more specific, or raise detection_thres above 0.50. Length A default run is 81 frames at 16 fps — about 5 seconds, taken from the start of your clip. For longer output raise these two together and keep them equal: • Load Video → frame_load_cap • Wan SCAIL To Video → length 161 ≈ 10 seconds, 321 ≈ 20 seconds. SCAIL-2 is trained at 81 frames, so longer runs cost more VRAM and the character may drift. To start later in the clip, set Load Video → skip_first_frames (frames at 16 fps — 160 skips 10 s). Running the same clip in 81-frame slices at 0, 81, 162, 243 and joining the files is the alternative to one long render; expect a visible seam at each join. Sound This workflow does not do sound. The render is always silent — SCAIL-2 generates picture only, and nothing in the graph carries audio through to the output. Add the soundtrack afterwards in a video editor, using your original clip as the audio source. Trying to attach it here is not worth it: the render is a short slice of your clip, so the audio would not line up anyway. Video format The upload button accepts .mp4, .webm, .mkv and .gif. H.264 MP4 is the safe choice. If a clip is rejected with "Invalid video file", re-encode it before uploading. A common cause is a movie-rip audio track (AC-3) that the pod's ffmpeg cannot decode: ``` ffmpeg -i input.mp4 -c:v libx264 -pix_fmt yuv420p -an clean.mp4 ``` -an drops the audio. Since this workflow keeps the original scene, re-encode from the highest-quality source you have — the output resolution is taken from this clip. Leave alone unless you know why • Replace mode is already on — the toggles on Create SCAIL-2 Colored Mask and Wan SCAIL To Video are true. Turning them off gives the animation behaviour instead (original background discarded). • Negative Prompt — a fixed quality filter, not a place to describe your video. • KSampler — steps 6, cfg 1.0, euler / simple, tuned for the distilled LoRA. Raising steps or cfg makes results worse, not better. • Output size comes from the video via Get Image from Batch, not from your reference image — so a portrait clip stays portrait no matter what you upload. • Every model this workflow needs is already installed on the pod.

ControlNet (depth) — klein 1. Load Image depth (red) — drop the photo whose structure you want to copy. A depth map is auto-extracted (Depth Anything V2). 2. Load Image reference (red) — drop the subject/style reference to place into that structure. 3. Prompt — describe the result (default refcontrol). 4. RefControl strength (yellow) — the Lora node. Higher = follow the depth structure more strictly. 5. Press Run — result in Save Image. Powered by the RefControl depth LoRA for klein 9B. The first depth run downloads the preprocessor weights (~1.3GB) once.

Face Swap — DarkBeast (klein 9B) 1. Target photo (red, top-left) — the picture whose face gets replaced. Ships with example.png so the graph runs out of the box. 2. Face to swap in (red, bottom-left) — the source face/identity to paste on. 3. Prompt — inside the Face Swap node; keep it simple, e.g. swap the face of the person with the reference face, keep pose, expression and lighting. 4. Steps = 5, CFG = 1 — DarkBeast is a distilled BFS model tuned for 5 steps / CFG 1. Do not raise them — higher values make it worse, not better. 5. Color Match (orange) regrades the result to the target's lighting so the swap blends in. Lower strength (or 0) to disable. 6. Press Run — the result is saved in Save Image. DarkBeast Klein 9b V2 BFS — face-swap-specialized klein 9B (safetensors, loaded via UNETLoader).

Image Edit — klein (multi-reference) 1. Load Image (red, group image 1) — drop the reference picture. It ships with example.png so the graph runs out of the box. 2. Prompt — describe the edit inside the Image Edit node. 3. Reference images toggle — switch image 2 … image 10 on to use more reference pictures. Each toggle enables one Load Image group on the left. 4. Color Match (orange) — regrades the result to image 1's lighting/colors so edits blend in. Lower strength (or 0) to disable. 5. Press Run — the result is saved in Save Image. FLUX.2 klein 9B (uncensored) — fast text → image + up to 10 reference images.

How to use 1. Prompt (orange) — type what you want to generate. 2. Two variants: Standard and Distilled (faster). Enable one and bypass the other with Ctrl-B. 3. Press Run — the result is saved in Save Image. FLUX.2 klein 9B (uncensored) — fast text → image.

How to use 1. Prompt Builder (red) — type your prompt in the Description field (plain language is fine). For layout control, open the Ideogram 4 editor and drag boxes to place objects/text in regions. 2. Resolution Selector (orange) — choose aspect ratio / size. 3. Press Run — the image appears in Save Image. "Image blocked by safety filter" comes from the model's own safety training, not ComfyUI.

How to use 1. Load Image — upload image 1 (required). Type the edit instruction in the Image Edit prompt field. 2. To combine pictures, enable Load Image 2 / 3 (right-click → Set Mode → Always) and upload. 3. Press Run — result in Save Image. Predefined example — reset every GPU start; use Workflows → Save As to keep your own copy.

How to use 1. Load Image (red) — drop the photo whose camera angle you want to change. 2. Qwen Multiangle Camera (red) — drag the 3D handle to set the angle, or pick a preset. The prompt is built for you. 3. Press Run — the re-angled image appears in Save Image. Powered by Qwen-Image-Edit-2511 + the multi-angle camera LoRA (4-step Lightning).

The quality of the style transfer depends largely on the quality of the RF inversion. These settings work well, but feel free to try other values.

How to use 1. Text to Image (red) — type your prompt in the text field, set width / height (and seed if you want). 2. Press Run — the image appears in Save Image. Sizes: 1:1 1328×1328 · 16:9 1664×928 · 9:16 928×1664 · 4:3 1472×1104 · 3:4 1104×1472 Predefined example — reset on every GPU start. Use Workflows → Save As to keep your own copy.

Upscale to 4K 1. Load Image (red) — drop any image (e.g. one you made with the Text-to-Image workflow). 2. Target size (yellow) — Scale to Total Pixels sets the working resolution. 4 MP ≈ 4K; raise/lower for your GPU. 3. Refine (yellow) — the KSampler re-renders detail at the new size. denoise ~0.35–0.45: higher = more new detail, lower = closer to the original. 4. Press Run — the upscaled image lands in Save Image. This is a single refine pass on an existing image. Generate first in the Text-to-Image workflow, then upscale here.

How to use 1. Load Image (red) — upload the image you want to edit. 2. Instruction — double-click the red Image Edit (Boogu) subgraph and type what to change in the prompt box. 3. Size (yellow) — output matches the input by default. Bypass Resize Image/Mask to keep the original size, or raise its megapixels (e.g. 4) for higher resolution — depends on your GPU. 4. Press Run — compare input vs result in Image Compare; the result is saved in Save Image. Boogu Edit is instruction-based image editing with strong bilingual (English / 中文) text rendering.

How to use 1. Prompt — double-click the red Text to Image (Boogu Turbo) subgraph and type your description in the prompt box. 2. Resolution (yellow) — pick aspect ratio / size in Resolution Selector. 3. Press Run — the image appears in Save Image. Boogu Turbo is a fast text-to-image model with strong bilingual (English / 中文) text rendering. For instruction-based image editing, open the Boogu Edit workflow.

How to use 1. Prompt — double-click the red Text to Image (Krea-2 Turbo) subgraph and type your description in Text String (User Prompt). 2. Resolution (orange) — pick aspect ratio / size in Resolution Selector. 3. Press Run — the image appears in Save Image. Prompt enhancement is on by default; it expands your prompt using the model's own text encoder (no extra model needed). Toggle prompt_enhance inside the subgraph to turn it off. Style LoRAs — set enable_lora? to true inside the subgraph, pick a krea2_* file in LoraLoaderModelOnly; the matching trigger word is added automatically. All 9 LoRAs are pre-installed.

Script → per-section SRT 1. Load audio (red, left) — upload the recording of your narration. 2. script (in the Align node) — paste your script; a blank line starts a new section. Each section becomes one SRT cue. 3. language — leave auto, or set ru / en / zh. 4. Press Run — WhisperX aligns your exact text to the speech and writes an SRT with one cue per section (exact start/end). A ⬇ Download SRT button appears when it finishes. Sections whose words aren't found in the audio (e.g. a line you skipped while reading) are dropped and logged. Runs best on a GPU tier.

AUDIO -> SRT (transcribe, keep timing). 1. Upload your source audio to 'Load source audio'. 2. On 'Audio -> SRT' pick the spoken language (auto / en / ru / zh). ENGINE is whisperx — Whisper core + phoneme forced-alignment; the tightest word-level timing and the best accuracy of the two engines. The Qwen3-TTS bundle additionally offers 'qwen3-asr' as an alternative engine; every other voice bundle is WhisperX-only. 3. Choose where the subtitle lines break: PAUSE_GAP (seconds) — a silence longer than this starts a new line. 1.5 is a natural sentence break; lower it for shorter, denser lines. MAX_CHARS — a line is also broken once it grows past this many characters. Set it to 0 to switch the character limit OFF and break on pauses only. 4. Run. The timestamped subtitles are saved to output/subs/transcript.srt and shown in the node. Then translate that SRT text (keep the timestamps unchanged) and feed it to the 'SRT -> Audio' workflow to get a dubbed track with the SAME timing.

SRT -> dubbed AUDIO, voice cloned, original timing preserved (Fish Audio S2 Pro). 1. Paste your TRANSLATED subtitles into 'Fish S2 SRT Dub' (keep the same timestamps as the source SRT, only the text is translated). 2. Upload a clean 10-30s clip of the speaker to 'Reference voice'. 3. Run. The output language is detected from the SRT text itself. Optional: press '🎙 Transcribe' on the dub node to preview the WhisperX transcription of your reference clip in 'ref_text' and fix it before Run. Left empty, the transcript is derived automatically. Knobs: fit_to_timing (on = lock each line into its SRT slot), max_stretch (cap before audio sounds sped-up).

TTS + VOICE CLONE (Fish Audio S2 Pro, 80+ languages). 1. Upload a clean 10-30s clip of the target speaker to 'Reference voice'. 2. 'ref_text' — the exact words spoken in that clip. LEAVE EMPTY to have it transcribed automatically (WhisperX); type it manually for maximum accuracy. 3. Type the text to speak into 'text' — the language is detected from the text itself. 4. Run, listen in 'Save audio'. Tip: leave 'Reference voice' unconnected to let the model pick a random voice. The S2 server starts at boot; the first request after boot may wait a bit while it warms up.

Script → per-section SRT 1. Load audio (red, left) — upload the recording of your narration. 2. script (in the Align node) — paste your script; a blank line starts a new section. Each section becomes one SRT cue. 3. language — leave auto, or set ru / en / zh. 4. Press Run — WhisperX aligns your exact text to the speech and writes an SRT with one cue per section (exact start/end). A ⬇ Download SRT button appears when it finishes. Sections whose words aren't found in the audio (e.g. a line you skipped while reading) are dropped and logged. Runs best on a GPU tier.

AUDIO -> SRT (transcribe, keep timing). 1. Upload your source audio to 'Load source audio'. 2. On 'Audio -> SRT' pick the spoken language (auto / en / ru / zh). ENGINE is whisperx — Whisper core + phoneme forced-alignment; the tightest word-level timing and the best accuracy of the two engines. The Qwen3-TTS bundle additionally offers 'qwen3-asr' as an alternative engine; every other voice bundle is WhisperX-only. 3. Choose where the subtitle lines break: PAUSE_GAP (seconds) — a silence longer than this starts a new line. 1.5 is a natural sentence break; lower it for shorter, denser lines. MAX_CHARS — a line is also broken once it grows past this many characters. Set it to 0 to switch the character limit OFF and break on pauses only. 4. Run. The timestamped subtitles are saved to output/subs/transcript.srt and shown in the node. Then translate that SRT text (keep the timestamps unchanged) and feed it to the 'SRT -> Audio' workflow to get a dubbed track with the SAME timing.

CHANGE VOICE (CosyVoice 3 native voice conversion) — keeps the words and the delivery, swaps only the timbre. 1. Upload the speech you want re-voiced (any length) to 'Source speech'. It is auto-split into <=25s chunks, re-voiced, and stitched back, so length is unlimited. 2. Upload a SHORT clean clip (3-30s) of the target speaker to 'Target voice'. Keep it SHORT (<=30s). 3. Run, listen in 'Save audio'. This is REAL voice conversion — no transcription step, so pauses, emphasis and pacing survive intact. To generate NEW speech from typed text instead, use the 'Voice Clone' workflow. First run downloads the CosyVoice model (~GB) — give it a few minutes. ⚠️ Predefined example — it resets to the original every GPU start; edits here are lost. Save under a NEW name (Save As) to keep your own copy; custom workflows persist between sessions.

SRT -> dubbed AUDIO, voice cloned, original timing preserved. 1. Paste your TRANSLATED subtitles into 'CosyVoice SRT Dub' (keep the same timestamps as the source SRT, only the text is translated). 2. Upload a SHORT clean clip (3-15s, no music) of the speaker to 'Reference voice' — CosyVoice clones this timbre. 3. Pick language (auto works; en/zh are reliable). 4. Run. Each line is synthesized and time-stretched to fit its slot, so the output lines up with your video. Knobs: fit_to_timing (on = lock to SRT timing), max_stretch (cap before audio sounds sped-up), speed. Languages: EN and ZH are solid. Other officially supported languages can be less reliable in this model — prefer the Qwen3-TTS bundle for those. First run downloads the CosyVoice model (~GB).

VOICE CLONE (CosyVoice 3 zero-shot). 1. Upload a SHORT clean clip (3-30s, no music) of the target speaker to 'Reference voice'. A reference clip is REQUIRED. 2. Type the text you want spoken into 'text' on the Voice Clone node. 3. Run, listen in 'Save audio'. First run downloads the CosyVoice model (~GB) — give it a few minutes. To re-voice EXISTING speech instead of generating new speech, use the 'Change Voice' workflow — it keeps the original words and delivery and only swaps the timbre. ⚠️ Predefined example — it resets to the original every GPU start; edits here are lost. Save under a NEW name (Save As) to keep your own copy; custom workflows persist between sessions.

Script → per-section SRT 1. Load audio (red, left) — upload the recording of your narration. 2. script (in the Align node) — paste your script; a blank line starts a new section. Each section becomes one SRT cue. 3. language — leave auto, or set ru / en / zh. 4. Press Run — WhisperX aligns your exact text to the speech and writes an SRT with one cue per section (exact start/end). A ⬇ Download SRT button appears when it finishes. Sections whose words aren't found in the audio (e.g. a line you skipped while reading) are dropped and logged. Runs best on a GPU tier.

AUDIO -> SRT (transcribe, keep timing). 1. Upload your source audio to 'Load source audio'. 2. On 'Audio -> SRT' pick the spoken language (auto / en / ru / zh). ENGINE is whisperx — Whisper core + phoneme forced-alignment; the tightest word-level timing and the best accuracy of the two engines. The Qwen3-TTS bundle additionally offers 'qwen3-asr' as an alternative engine; every other voice bundle is WhisperX-only. 3. Choose where the subtitle lines break: PAUSE_GAP (seconds) — a silence longer than this starts a new line. 1.5 is a natural sentence break; lower it for shorter, denser lines. MAX_CHARS — a line is also broken once it grows past this many characters. Set it to 0 to switch the character limit OFF and break on pauses only. 4. Run. The timestamped subtitles are saved to output/subs/transcript.srt and shown in the node. Then translate that SRT text (keep the timestamps unchanged) and feed it to the 'SRT -> Audio' workflow to get a dubbed track with the SAME timing.

RE-DUB VOICE (Qwen3-TTS pipeline: transcribe -> re-speak, timing preserved). This is NOT voice conversion. Qwen3-TTS has no native VC, so this workflow chains two steps: the source audio is transcribed with timestamps (Audio -> SRT), then every line is re-spoken from scratch by the cloned TARGET voice at its original timestamp (SRT Dub). 1. Upload the recording you want re-dubbed to 'Source audio'. 2. Upload a SHORT clean clip (3-15s, no music) of the TARGET voice to 'Target voice'. Leave 'ref_text' empty — it is transcribed automatically. 3. On the dub node pick the LANGUAGE of the source speech and run. Line timing is preserved, but the words are re-generated, so intonation, pauses and emphasis inside each line are the model's, not the original speaker's. Two side effects worth knowing: a transcription error becomes a wrong word in the output, and any non-speech audio is dropped. For REAL voice conversion — same words, same delivery, only the timbre swapped — use the CosyVoice 3 bundle (the change-voice pick) or Chatterbox Multilingual. Both do it natively with no transcription step.

SRT -> dubbed AUDIO, voice cloned, original timing preserved (Qwen3-TTS). 1. Paste your TRANSLATED subtitles into 'Qwen3-TTS SRT Dub' (keep the same timestamps as the source SRT, only the text is translated). 2. Upload a SHORT clean clip (3-15s, no music) of the speaker to 'Reference voice'. 3. 'ref_text' — the exact words spoken in that clip. LEAVE EMPTY to have it transcribed automatically (WhisperX); type it manually for maximum accuracy. 4. Pick the language of the TRANSLATED text and run. Knobs: fit_to_timing (on = lock each line into its SRT slot), max_stretch (cap before audio sounds sped-up). Each line is synthesized with the cloned voice and placed at its SRT timestamp, so the output lines up with your video.

VOICE CLONE (Qwen3-TTS 1.7B Base). 1. Upload a SHORT clean clip (3-15s, no music) of the target speaker to 'Reference voice'. 2. 'ref_text' — the exact words spoken in that clip. LEAVE EMPTY to have it transcribed automatically (WhisperX); type it manually for maximum accuracy. 3. Type the text you want spoken into 'text' and pick its language. 4. Run, listen in 'Save audio'. First run loads the model (pre-baked at boot, a few seconds).

VOICE DESIGN (Qwen3-TTS 1.7B VoiceDesign). No reference audio needed — describe the voice you want in plain language. 1. Type the text to speak into 'text'. 2. Describe the voice in 'instruct' (gender, age, mood, pace, accent — e.g. 'A raspy old pirate, slow and theatrical'). 3. Pick the language and run. Tip: to REUSE a designed voice, save its output and feed it into the Voice Clone workflow as the reference sample.

Script → per-section SRT 1. Load audio (red, left) — upload the recording of your narration. 2. script (in the Align node) — paste your script; a blank line starts a new section. Each section becomes one SRT cue. 3. language — leave auto, or set ru / en / zh. 4. Press Run — WhisperX aligns your exact text to the speech and writes an SRT with one cue per section (exact start/end). A ⬇ Download SRT button appears when it finishes. Sections whose words aren't found in the audio (e.g. a line you skipped while reading) are dropped and logged. Runs best on a GPU tier.

AUDIO -> SRT (transcribe, keep timing). 1. Upload your source audio to 'Load source audio'. 2. On 'Audio -> SRT' pick the spoken language (auto / en / ru / zh). ENGINE is whisperx — Whisper core + phoneme forced-alignment; the tightest word-level timing and the best accuracy of the two engines. The Qwen3-TTS bundle additionally offers 'qwen3-asr' as an alternative engine; every other voice bundle is WhisperX-only. 3. Choose where the subtitle lines break: PAUSE_GAP (seconds) — a silence longer than this starts a new line. 1.5 is a natural sentence break; lower it for shorter, denser lines. MAX_CHARS — a line is also broken once it grows past this many characters. Set it to 0 to switch the character limit OFF and break on pauses only. 4. Run. The timestamped subtitles are saved to output/subs/transcript.srt and shown in the node. Then translate that SRT text (keep the timestamps unchanged) and feed it to the 'SRT -> Audio' workflow to get a dubbed track with the SAME timing.

VOICE CONVERSION (Chatterbox VC). Replaces the VOICE in a recording while keeping the words, pacing and intonation of the original. 1. Upload the recording you want to convert to 'Source audio'. 2. Upload a SHORT clean clip (3-15s, no music) of the TARGET voice to 'Target voice'. 3. Run, listen in 'Save audio'. First run loads the VC model (pre-baked at boot, a few seconds).

SRT -> dubbed AUDIO, voice cloned, original timing preserved (Chatterbox Multilingual v3). 1. Paste your TRANSLATED subtitles into 'Chatterbox SRT Dub' (keep the same timestamps as the source SRT, only the text is translated). 2. Upload a SHORT clean clip (3-15s, no music) of the speaker to 'Reference voice'. No transcript needed. 3. Pick the language code of the TRANSLATED text (en / ru / zh / ...) and run. Knobs: fit_to_timing (on = lock each line into its SRT slot), max_stretch (cap before audio sounds sped-up). Each line is synthesized with the cloned voice and placed at its SRT timestamp, so the output lines up with your video.

TTS + VOICE CLONE (Chatterbox Multilingual v3, 23 languages). 1. Upload a SHORT clean clip (3-15s, no music) of the target speaker to 'Reference voice'. No transcript needed. 2. Type the text to speak into 'text' and pick its language code (en / ru / zh / ...). 3. Run, listen in 'Save audio'. Tip: leave 'Reference voice' unconnected to use the model's default voice. First run loads the model (pre-baked at boot, a few seconds).