Repository navigation
Releases
H3 Editor 0.13.0
-
Experimental Sprite Draft adds a directional motion guide to independent shots. Choose a cardinal facing and an eye-level, RPG or overhead view. Preview the loop at game size and download transparent illustrated and indexed PNG strips. Finish draft keeps the selected take’s seed, loop timing, scale and pivot.
-
Sprite Draft can generate a planted idle loop at 6 fps. Auto cutout uses an installed BiRefNet model on the selected poses, removes the backdrop colour from soft edges and filters coverage at game size. The background key remains available. Existing Sprite Draft takes export at 64, 128 and 256 px per cell from the saved video, with the original poses, timing and ground pivot. Both PNG styles are cached per size.
-
Crop or reframe a placement, export a marked timeline range, and mix sound with fades and volume points. The editor and agent API share the same project edits.
-
Choose additional aspect ratios, calculate a resolution by longest side or megapixels, and remove selected takes from a clip. Removed takes can be restored from history.
-
Add experimental mid-frame image guides in the Shot inspector or agent API. Upload a picture or capture a frame from the viewer. The guide frame is measured in the source shot.
-
Try the optional VOSR2 2× upscaler on imported footage and final takes. Choosing it explicitly produces a separate take and preserves the source audio. Fine texture can flicker between frames.
-
Open contextual actions with right-click or Shift+F10 on the bin, viewer, timeline, inspector and dialogs. Extend and Bridge appear together on timeline clips. Compact windows can switch between Clips and cast, Viewer and Inspector.
-
Short footage imports now use a poster frame inside the clip. Shot duration errors show the value the project kept.
-
Edit a 2–15 second range with a prompt and compare two takes with a synchronized swipe. The clips panel also offers a grid view.
-
Edit track volume, fades and volume points on the waveform. Drag style markers directly on the Look lane.
-
Experimental motion transfer uses a source performance and a character reference to generate a new take with the source audio. It requires the pose and ControlNet models listed in the README.
H3 Editor 0.12.0
H3 Editor 0.12.0 adds reusable looks and experimental style changes within a single generated shot. It also adds playback ranges, a spectrogram, and more control over takes and final renders.
Looks and style transitions
- Style library: Start with one of eight built-in looks or save your own. Saved looks live outside project folders and are available in other projects on the same ComfyUI installation.
- Style markers (experimental): Set different looks at different frames of a generated clip. H3 blends their conditioning over the shot. The attached three-style demo goes from graphite to layered cut paper to clay in one continuous shot. See how to reproduce it and what to expect. Try a small draft first: every distinct style adds a model evaluation per sampling step.
Editing and rendering
- Loop playback: Use the toggle beside Play to switch between playing once and looping. Set In and Out to loop a range in the cut or a clip preview.
- Sound lanes: Switch the timeline between waveform and spectrogram.
- Independent takes: Expand a clip in the Clips panel, then drag a specific take onto the picture track or append it as its own clip. You can trim it without changing which take the original clip uses. (#11)
- Batch finals: Queue eligible final renders together and cancel only finals or every render in the current project. (#9)
- More finish sizes: Finish a draft from its saved latents at Large, Medium, Small or a custom size, as well as Full. A finish preserves the draft's style markers. (#8, #10)
Update H3 Editor and restart ComfyUI to register the new H3 Style Guider node. The demo was rendered at 1344 × 768 and retains all 175 frames of the original draft. We also exported a cut joining the first part of the draft to the second part of its finish.
Full changelog: v0.11.0...v0.12.0
H3 Editor 0.11.0
H3 Editor 0.11.0 turns the draft you like into a final, the same shot at full size, in a fraction of a final's time.
Highlights
- Finish draft. Make the final from a draft you already like, instead of rendering a new one and hoping it comes out the same.
- Faster finals. Finishing runs only the final's last steps: a 3.8 s vertical shot finished in 88 seconds, where a final of that length took about four minutes on the same GPU.
- A final no longer looks stuck. The refine pass has its own progress labels.
Finish draft
A final has always been a new render: it samples at a larger size, and the same seed at a different size draws a different picture. So a draft you liked could not be brought back at final size, even with its seed fixed.
Finish draft starts from the draft itself. Every draft keeps the latents it was sampled as, beside its video. Finishing loads them, enlarges them with the latent upscaler, and redraws them at full size in place of a new first pass.
| Where | Finish draft under a clip whose shown take is a draft, and Finish this draft in its right-click menu |
| What you get | A new final take with the draft's framing, motion, faces and sound, at full detail |
| Time | Only the final's refine steps run |
| Works best with | Draft size: Medium in the Project tab, exactly half the full size |
| Needs | Final size: Full, and a draft rendered with 0.8.0 or later |
| Refused when | The shown take is not a draft, or the clip's length changed since the draft |
The new take records which draft it was finished from. Render final still renders a fresh take, for when you want a new one.
Progress during a final
A final's second pass enlarges the picture and redraws it at full size. It runs after the sound is decoded and, on smaller cards, takes longer than the first pass, but the editor had no label for it, so "Decoding the sound" stayed up for minutes. The queue now shows:
| Stage | What runs |
|---|---|
| Enlarging | The latent upscaler brings the picture to full size |
| Refining at full size | The last steps redrawn at full size, with step progress |
| Decoding the full-size picture | The full-size frames are decoded |
If that pass is too slow on your card, set Final size to Large (no second pass) or a size of your own.
For agents
| What | Where |
|---|---|
| Finish a draft | render with "quality": "finish" on a clip whose active take is a draft |
| The draft it came from | finished on the new take |
| H3 Load Take | A node that reads a take's saved latents, picture and sound, as an H3 latent |
Updating
Update the pack and restart ComfyUI, so the new node registers. Projects made with 0.10.0 need no change.
How it was tested
Every case was finished from a draft with Finish draft, then compared frame by frame with its draft and run through the cut report.
| Case | Draft | Finished in | Result |
|---|---|---|---|
| A single shot | Medium, 544×960, 3.8 s | 88 s | Same framing, motion and expression in every frame, with full-size detail in skin, hair and hands |
| A single shot from a Small draft | Small, 352×640, 3.8 s | 91 s | The same, from a draft enlarged about three times |
| An extension | Medium, 5.9 s take | 166 s | Matches its draft; the join with the shot it extends scores 97, continuous |
| A bridge | Medium, 6.4 s take | 267 s | Matches its draft; both its joins score 96 and 98, continuous |
The progress labels were watched in a browser while the Small draft finished from the editor's own button: Preparing the full size, Enlarging, Refining at full size with its steps, Decoding the full-size picture, Saving, Adding the take.
A finish keeps what its draft has, mistakes included: the bridge's draft cut from the café to the flower stall instead of moving through the door, and the finished bridge cuts at the same frame. Watch a draft before you finish it.
Times are on a 96 GB GPU; a smaller card spends longer in the refine pass.
H3 Editor 0.10.0
H3 Editor 0.10.0 fetches the models for you, keeps long shots together, plays the cut frame-exact, and checks every clip's lips and room.
Highlights
- Get the models. When the default models are missing, the editor lists what it needs, with sizes and licenses, and downloads them into the right folders with one button.
- Long shots hold together. Extensions, bridges and merges no longer drift harsher, shinier or more orange with each clip, and long finals refine in context windows.
- Two new checks. Checks now says when a clip's lips don't follow its sound, and when its room changes under the person.
- The cut plays frame-exact. No more micro jumps at cuts in the viewer, and a paused frame is the frame you asked for.
- Volume per clip, and a volume for the viewer.
- Fixes for two reported issues: no more
Pin error.floods on cards that offload HyperFlow (#4), and an option to free GPU memory for a local writing model (#6).
Get the models
On a ComfyUI without H3's model files, the editor opens a panel that names every file HyperFlow H3 Flash needs to render drafts and finals, with each one's size and license, about 54.6 GB in all. One button downloads them from Hugging Face into ComfyUI's own models folders, with progress, speed and time left for each file and for the whole.
| Where | A panel that opens by itself when files are missing, and Get the models in the Project tab while any file is missing |
| What | The HyperFlow H3 Flash model, the text encoder, the video and audio VAEs, and the latent upscaler for finals |
| Safe | Only files on the editor's own list, only from huggingface.co over https. A file comes in under a .part name and is renamed only once its sha256 matches the published file. A file already in place, or another precision of it, is never replaced |
| Stop and resume | Stop keeps what came; the next download carries on from there |
| Not included | The Realism Slider LoRA: Civitai asks for a sign-in, so the panel links to its page |
On a ComfyUI that already has every file, nothing changes: the panel stays closed.
Long shots hold together
A long shot made of a head clip and its extensions used to drift: each extension held its source's frames and added a small lean of its own, and the leans added up.
| What changed | Effect |
|---|---|
| Held frames far from the join are partly redrawn, falling to nothing at the join | A head clip and four extensions on HyperFlow came out 4, 16, 28 and 31% harsher than the head; now -3, 3, 6 and 3%, with joins scoring the same |
| Linked clips run the Realism Slider at half the project's strength (0.6 at the usual 1.2) | A sung head clip and four extensions no longer pile texture up (it was 5, 18, 33 and 46% harsher); the chain stays within 1% of the head by the third clip |
| An extension, bridge or merge keeps its source clip's model and LoRAs | Extending a clip set to another model no longer changes the look where the two meet |
| Finals longer than 10.4 s refine in context windows of 124 frames | A 20.75 s sung take came out the same as one refined whole, in 15 minutes instead of 23 |
Clips rendered before keep what they have; render a linked clip again to get the fixes. Long finals need the core MiniMaxH3ContextWindows node; without it they refine whole as before.
Checks
| Check | What it finds | Needs |
|---|---|---|
| Lips | Clips whose lips run ahead of their sound, behind it, or don't follow it, with the offset. It uses SyncNet, written into the pack from its MIT code. | SyncNet's weights and a small face detector in models/syncnet |
| Place | A generated clip whose background, the person taken out, is no longer what it was: a keyboard turning into a monitor, a chain of extensions drifting into another room. It follows the camera's move. | MODNet's portrait matting file in models/matting |
| Sound runs out | A take whose sound stops half a second or more before its picture, from which frame | Nothing |
The model files for Lips and Place are not downloaded for you yet; the models page says where they come from. Without them, those two checks stay quiet.
Playback
| Before | Now |
|---|---|
| The viewer switched clips a moment late: at most cuts it showed a frame from past the out-point, then the next clip's frame from before its in-point, and some cuts froze briefly | A cut is timed from the last frame the outgoing clip put on screen, and the next clip stays hidden until it shows its in-point. On a 43 s reel of 14 clips, all 1031 frames now play in order, where 9 frames from outside the cut used to show |
| Stepping frame by frame often showed the frame before the one asked for | A paused frame is the frame you asked for |
Sound
- A volume for each clip. The Timeline tab of a clip has a Volume slider: 1 as rendered, 0 silent, up to 2 louder. It belongs to the clip's place on the timeline, and the export applies it.
- A volume for the viewer. A speaker button and slider beside the play controls set how loud the editor plays. It changes nothing in the export.
Writing
- Room on the card for a local writing model (#6). The writer's settings have Free ComfyUI's GPU memory before each request, for the OpenAI-compatible and Ollama providers, off by default. ComfyUI unloads its models before each rewrite, unless a render runs or waits, and the next render loads them again. Have the model server unload its own model when idle (a
ttlin llama-swap,keep_alivein Ollama) so renders get the card back. - Sung words stay out of the action. When someone sings, the words go in the dialogue and the action describes the singing; quoting the lyric in the action gave a looser lip sync in 5 of 6 seeds.
Fixed
- No more
Pin error.warnings (#4). On a card too small for HyperFlow, ComfyUI tried to pin every layer it kept in system memory, where it sat in ComfyUI's read-only mapping of the model file. CUDA cannot pin that memory, on Windows or Linux, so every layer printed the warning. HyperFlow now skips those attempts, and each one also waited on the GPU, so steps can come out a little faster. - The cuts report stays up when a check cannot run, instead of failing whole.
For agents
| What | Where |
|---|---|
gain on a timeline item |
update_item fields |
lips and scene in the cuts report |
GET /project/{name}/cuts; h3.py cuts prints a lips: line |
context_noise |
H3 Take node input |
| Model setup | GET /h3take/api/models/setup says what is missing; POST /h3take/api/models/setup/download starts the download and .../cancel stops it. 409 while one runs, 507 when a disk is short |
Updating
Update the pack and restart ComfyUI. Projects made with 0.9.x need no change.
Not tested yet
- Get the models was tested with unit tests, a real resumed download of one small file from Hugging Face, and the panel against a stand-in server with small files in place of the real ones. Not yet run: a full 54.6 GB download on a ComfyUI missing the files, a draft rendered right after it, the out-of-space refusal on a real full disk, filesystems without hard links (some Docker bind mounts), and installs whose
extra_model_pathsput the models elsewhere. - No more
Pin error.was tested on the cause itself (CUDA refusing to pin a read-only file mapping, on Windows and Linux) but not on a card small enough to offload HyperFlow. - Playback was measured in Chrome on one 14-clip reel without transitions or speed changes.
H3 Editor 0.9.1
H3 Editor 0.9.1 lets a project render at any size you type, for drafts and for finals.
Highlights
- Your own size. Draft size and Final size in the Project tab now end with Your own size: a width and a height from 256 to 2048 pixels.
- A 0.6 megapixel size out of the box. The boxes start at 1024×576, or 576×1024 in a vertical project, between Medium and Large.
- Any shape. A size such as 1344×576 works in a 16:9 project.
Your own size
Until now a project picked from five named sizes. Pick Your own size and two boxes appear under the menu, with the megapixels beside them.
| Range | 256 to 2048 pixels on each side |
| Steps | Multiples of 32, which H3's latent needs. What you type moves to the nearest one: 1280 by 720 becomes 1280 by 736. |
| Starts at | 1024×576 in a wide project, 576×1024 in a vertical one, 768×768 in a square one |
| For | Drafts and finals |
| A final at this size | Comes out as sampled, with no enlarging pass, like Large |
What to expect:
- The shape can differ from the project's format. The editor says so under the boxes. The cut and the export take the shape of the largest picture on the timeline, and every other clip fills it.
- H3 was trained near one megapixel. Far above or below that the picture can suffer, so look at a draft before a long render.
- Takes already rendered keep their size. A clip that extends or bridges a take of another size holds its frames from the video, as it does after any change of size.
- Every render is 24 frames a second, whatever its size. H3 makes picture and sound together at that rate and has no setting for another.
Measured at 1024×576 on MiniMax H3 Turbo, 6 steps: a 3.8 second final took 32 seconds once the models were loaded.
For agents
| What | Where |
|---|---|
| A size of the project's own | set_project sizes, written "1024x576": {"draft": "832x480", "final": "1024x576"}. A size off the grid or out of range is refused, with the rule in the message. |
own under each format in sizes |
GET /projects: step, min and max for such a size, and first, the one to start from. |
name in the summary's sizes |
The named size, or the size as written. |
Updating
Update the pack and restart ComfyUI. Projects made with 0.9.0 need no change.
H3 Editor 0.9.0
H3 Editor 0.9.0 lets a project render with speed-up nodes from your other packs: sparse attention, low-memory attention, a step cache, a compile node. It also closes two gaps in how project zips are opened.
Highlights
- Model patches. Any installed node that takes the model and returns it can go on every render of a project.
- The settings are the node's own. The editor reads them from the node's definition, so a pack released tomorrow works without a change to the editor.
- Project zips open more carefully. A zip made by hand can no longer bring a revision history, and its patches open switched off.
Model patches
Other packs ship nodes that change how the model runs without being a model: sparse or low-memory attention, a cache that skips steps, a compile node. In a workflow you put one between the loader and the sampler. The Model patches section of the Project tab does the same for every render of a project.
| In the section | What it does |
|---|---|
| Add a patch | Lists the installed nodes that take one model and plain values and return one model. The ones for H3, MiniMax or attention come first; type to search by name, pack or category. |
| The settings in a patch's row | The node's own: numbers with their range, switches, menus and text. The button under the name shows what you changed from the defaults, and Reset goes back to them. |
| The tick box | Renders without the patch and keeps it in the list. |
| The arrows | Move a patch up or down. Patches go on the model in the order listed. |
| The bin | Removes the patch. |
How they are applied:
- Patches sit after the LoRAs, on drafts and on both passes of a final.
- They are set for the project, up to 8. A clip cannot have its own.
- A take records the patches it was made with.
- A patch whose pack is not installed shows Not installed and is left out of renders, so the project still opens and renders on a machine without that pack.
- A project opened from a zip has its patches unticked until you tick them: whoever made the zip chose the nodes and their settings.
Does a patch help?
The editor cannot know. Some patches speed a render up at a cost to the picture, some only save memory, and some do not work with every checkpoint. Render the same clip with the same seed with the patch and without, compare the times, and look at both. For clips that join, h3.py cuts scores the cuts of each version.
One was measured for this release, Veda Sparse Attention at 90% sparsity, on MiniMax H3 Turbo at 6 steps with the same seeds:
| Render | Without | With | |
|---|---|---|---|
| Draft, 640×352, 3.8 seconds | 22 to 24 s | 51 to 58 s | 2.3 times slower |
| Final, 3.8 seconds | 226 and 235 s | 186 and 173 s | 1.3 times faster |
| Final of a bridge, 8 seconds | 661 s | 288 s | 2.3 times faster |
The bridge is one run each way. Its two cuts scored the same with the patch and without: 98 of 100 for the picture at both, and 94 against 95 for the sound into the next clip.
The time is not the whole answer. All three 3.8 second finals made with the patch have a small jump 5 frames before their end, at the last step of the picture latent. The finals made without it and the drafts do not. The editor's Checks list it as a change of picture inside the take, and h3.py cuts reports it under inside.
What that says about patches in general:
- Compare at the size you mean to use. A patch with a fixed cost at every step loses on a short draft and wins on a long final. Untick it for drafts.
- Read the checks of a take made with a patch, and look at its end. A faster render can differ in a few frames only.
- A patch may want a certain checkpoint. That node accepts plain MiniMax H3 checkpoints and refuses HyperFlow H3 Flash, the editor's default model. When a render is refused, the message names the patch.
Project zips
| Fixed | What happened |
|---|---|
| A zip can no longer bring a revision history | A zip made by hand could carry files that landed in the new project's history, its previews or its project file, and restoring such a revision brought back whatever it held. Opening a zip now passes over everything that Download as zip never packs, and the project starts with one revision. |
| Paths that Windows reads as another file are refused | A name with a colon, a control character, one of < > " | ? *, or a device name such as NUL. The editor writes no such names. |
Zips made by the editor open as before.
For agents
| What | Where |
|---|---|
patches in set_project |
[{"node": id, "inputs": {name: value}, "on": true}]. Replaces the whole list; inputs names only what differs from the node's defaults; [] clears it. |
patches in GET /projects |
Every installed node a project can use, with its title, pack, category and controls: each input's name, kind, default, range or options, and the node's own tooltip. |
patches_off in the reply to POST /projects/open |
The patches a zip had on, opened switched off. Tell the user which they are, and switch one on only when asked. |
| Limits | A patch takes up to 64 inputs and text of up to 4000 characters. A render is refused, with the input's name, when a value is outside what the node accepts. |
Updating
Update the pack and restart ComfyUI. Projects made with 0.8 need no change, and nothing renders differently until you add a patch.
H3 Editor 0.8.0
H3 Editor 0.8.0 is about what happens where two clips meet. Joins hold a take's own latents and cut where the two files match, bridges land on the next clip, and the cuts report now listens to a cut and watches how things move across it.
Highlights
- Joins hold a take's own latents, from any cut, with its sound. A chain of extensions no longer gets darker and softer with every join.
- Joins cut where the two clips match, 9 frames inside the frames both files show.
- Finals join finals the same way, in the refine pass too.
- A bridge lands on the next clip. It used to cut to it.
- The cuts report listens and watches. Every cut gets a sound score, the camera's move across it, and what stops or starts moving in the frame. The editor lists what it finds.
- The cast catalog. Save a cast member once and bring them into any project.
- A project list you can see, with a frame of each project.
Joins
A clip made with Extend forward, Extend back or Bridge to next holds frames of the clip it joins. Until now those frames were read back from the source's video file and encoded again, except at a take's two ends. Each trip through the video decoder and encoder cost a little: over a chain of extensions the held frames came out 1.2 to 2.7 percent darker per join.
A take now saves all of its latents beside its video, and a linked clip holds them as they are.
| What is held | When it is the take's own latents | Otherwise |
|---|---|---|
| Picture | The cut is on a multiple of 17 frames, the grid of H3's picture latent. A whole clip always is. | Read from the video, as before |
| Sound | The cut is also on a multiple of 3 frames, as the end of every extension is. | Encoded from the file, as before |
| A final's refine pass | The source is a final too, refined at the same size. | Read from the video, as before |
Extending from a trimmed clip moves the cut onto the grid by 16 frames at most (two thirds of a second), and the clips after it close up. A whole clip is already on the grid. Imported footage keeps its cut, and Merge with next still holds its frames from the video.
The file beside a take is about 1.6 MB for a 5 second draft and about 22 MB for a 5 second final.
Where the cut sits
Both files show the held frames, and the timeline has to cut from one to the other somewhere. It used to cut at their edge, which is the worst place: a video decoder draws the last frames of a file with nothing after them and the first with nothing before them, and the held sound is softened next to the new material. New extensions, bridges and merges cut 9 frames inside the held frames. The source stops 9 frames earlier and the new clip plays those 9 itself, so the scene is as long as before.
| Measured | At the edge | 9 frames inside |
|---|---|---|
| Picture at the end of a file, against the same frames in the other file | 1.5 dB under the rest on the last frame, 5 dB at worst | Level with the rest |
| Picture at the start of a file | 6 to 9 dB under on the first frame | Level with the rest |
| Step in the sound at the cut, against the steps around it | 7 to 22 times | Under 3 times |
Joins made before this release keep their cut. Render a linked clip again after trimming its source and it is joined at the new cut.
Bridges
A bridge used to hold the next clip's first 39 frames at its end, as it holds the first clip's last 39 at its start. H3 treats a run that long from another shot as another shot: in 8 of 9 test pairs it followed the prompt and then cut to the next clip a few frames early. Changing the prompt did not help.
Bridge to next now holds a short run of the next clip, 22 frames with their sound, and the timeline cuts 9 frames inside it. That is enough to say where things are and how they move, and too little for H3 to take for another shot.
| At the cut into the next clip | One pinned frame (tried first) | 22 frames held (now) |
|---|---|---|
| Picture, of 100 | 71 to 97 | 97 to 99 |
| Sound, of 100 | 0 to 90, with drops of up to 36 dB | 78 to 97 |
| What moves in the frame | Stopped at the cut in one pair | Carried across every time |
Two test pairs, three seeds each, drafts.
| Bridge to next | Bridge into its movement | |
|---|---|---|
| Holds of the next clip | 22 frames | 39 frames, as before |
| Use it for | A change of framing or of place | A next clip that starts in mid-movement |
| Default length | 6.4 seconds | 4.25 seconds |
What to expect:
- A bridge needs a way from one picture to the other. It works best when the two shots share a place, or when the move passes through something that hides the change: into the dark of a cup, through a window. Two shots with nothing in common still cut.
- What happens on the way depends on the seed. The same bridge may push in, dissolve or cut. Look at the take, and render it again when it does not move as you asked.
- A move needs room. At 4.25 seconds a push-in from a wide shot to a close-up never came out as one shot in testing, so a new bridge is 6.4 seconds unless you set its seconds.
- The held sound can step a little at the cut. It is encoded from the next clip's file, because 22 frames is not a whole number of sound steps.
Bridges made before this release keep their way of arriving.
The cuts report
The report already scored the picture at every cut. It now scores four more things, and the editor lists what they find under Checks in the top bar.
| Score | What it catches | How |
|---|---|---|
| Sound at a cut | The level jumping, a swell chopped off, sound into silence, a click | Loudness on each side after ITU-R BS.1770, against how much the sound moves by itself around the cut; the change in the spectrum's shape; the step in the wave |
| Camera move at a cut | A push-in or pan that stops dead, starts or reverses at the cut | The optical flow of the whole frame (Lucas and Kanade's method, one pan and one zoom per frame) |
| Movement in the frame | Something that moves before the cut and is still after it: rain, a wave, a beam of light | What is left in each of 72 parts of the frame once the camera's move is taken out |
| Sound inside a take | A sound that is cut off, a hole in the sound, a take that ends 20 dB louder or quieter than it began | The level along the take. A sound that starts at once is left alone: words and waves do that |
Each score is judged against what the sound or the picture does by itself around the cut, so a handheld shot or a swelling sea is not flagged for being itself. A cut to a new shot has its camera move and movement described, not scored.
python joins.py PROJECT_FOLDER is a separate measure for joins: it compares the two copies of the frames every join holds, latents, picture and sound.
The cast catalog
| Action | Where |
|---|---|
| Keep a cast member for other projects | Save beside Catalog on the cast member |
| Bring one into a project | Add, From the catalog. The project gets its own copy of their pictures and voice. |
| Pick up a change | After a change is saved to the catalog, other projects show Newer in the catalog and an Update button. |
| Update every project at once | From the catalog. Copies changed in their own project are left alone. |
The project list
The project menu in the top bar is now a list of the editor's own: each project with a frame of its first rendered clip, its shape, how many clips are on its timeline, how long it runs and when it last changed. Type to narrow it by name, move with the arrow keys and open with Enter.
LoRA starts
H3 samples with a shift that keeps most of a render near the start of the scale, so a start of 0.75 was already four fifths of the way in. The presets moved to match.
| Preset | Was | Now |
|---|---|---|
| After framing | 0.75 | 0.93, about halfway |
| Details only | 0.4 | 0.7 |
| The Realism Slider on finals | 0.75 | 0.93: the last four of HyperFlow's six steps, where it used to work on the last three |
A start you set yourself stays as it is.
Fixed
- The side panel's tabs no longer make the panel scroll sideways at its usual width.
- Space plays again after you pick from a menu and then scrub the timeline.
For agents
| What | Where |
|---|---|
sound, movement and parts on each cut |
GET /project/{name}/cuts: what the cut sounds like, how the camera's move carries across it, and what stops or starts moving in the frame. Each has kind, score and notes. |
sound_inside and arrivals |
The same report: sound faults inside a take, and whether a bridge pinned to a frame reached it. worst counts all of these. |
"arrive" on bridge |
"hold" holds 39 frames of the next clip, "pin" pins its first frame alone. Left out, the bridge holds 22 frames, and its link says "arrive": "short". |
handover on a link |
How many frames inside the held frames the cut sits: 9 on new links. |
save_cast_to_catalog, add_cast_from_catalog, update_cast_from_catalog |
Operations for the cast catalog. GET /projects lists the entries under catalog, and each project's facts under cards. |
end_hold_frames on H3 Take, final on H3 Save Hold |
For your own workflows: a bridge that holds a short run of the second clip, and a refine pass saved beside the first. |
python listen.py a.mp4 b.mp4, python joins.py PROJECT_FOLDER --json |
The sound score and the join measurements from the command line. |
Updating
Update the pack and restart ComfyUI. Projects made with 0.7 need no change.
Takes rendered before this release keep their old file and are held from their two ends, and finals from their video in the refine pass. To get exact joins on an existing cut, render the source clip again, then the clip that joins it.
H3 Editor 0.7.0
H3 Editor 0.7.0 lets H3 run on cards that cannot hold its 32B text encoder, and makes the side panel shorter and easier to read.
Highlights
- A smaller text encoder. A Qwen3-VL 4B or 8B can stand in for H3's 32B text encoder through a projection, built into the pack: about 4.6 GB instead of 15.7 GB with the 4B. No other custom node is needed.
- A shorter side panel. Settings you set once moved into dialogs, short fields put their label beside them, and the tabs, sections and small actions have icons. At 1024 × 768 the Project tab scrolls about a quarter less, the Shot tab about a fifth less, and a cast member no longer scrolls at all.
- A LoRA's start reads as a timeline, with presets for the usual starts.
A smaller text encoder
This is the method of ComfyUI-ClipProj (MIT), ported into the pack with credit in NOTICE.
| File | Folder | Size |
|---|---|---|
qwen3vl_4b_int8_convrot.safetensors, or another precision of the 4B |
models/text_encoders |
4.5 GB |
mmh3-4b-ClipProj-v3.1.safetensors |
models/clip_projections |
26 MB |
With both installed, the 4B shows in the Text encoder menu: in the Project tab, press Edit beside Model files. With more than one projection for its size, a Projection menu picks between them, and the render takes v3.1 by default. An 8B works the same way with its own projection. The H3 Small Text Encoder node does the same in your own workflows.
What it costs: the small model knows less than the 32B. In our test the 4B followed the prompt and kept the cast member's face as the 32B did, but said one word of English dialogue not quite right, where the 32B said it cleanly. The bf16 4B and the larger v3.1-mlp projection did no better. Use it when the 32B does not fit, and listen to every line it speaks. Details are in models and quality.
The side panel
| Where | What changed |
|---|---|
| Project tab | A line at the top says what the project holds. Then Look, Format (with Language beside it and both sizes on one row) and Model. |
| Model files | The text encoder, video VAE and projection moved into the Models dialog. The panel shows one line with what renders, and an Edit button. |
| Shot tab | Rewrite is a small button in the label of What happens, with Camera and Seconds right under it. The first and last frames take one row of small pictures, with icon buttons to use the playhead frame, upload or remove. |
| Model and LoRAs | A clip's own model and LoRAs share one section. Its LoRAs, with their strengths and starts, are set in a dialog. |
| Cast member | The panel keeps the picture, the description and the name. Pictures, voice and RefMod are edited in a dialog with two columns. |
| Takes | Keep seed and Restore prompt have icons, and Delete take is a trash button at the end of the row. |
| Everywhere | Folding sections stay open or closed as you left them, in this browser. The Timeline tab shows whenever the clip is on the timeline, however you selected it. In a narrow panel, the tabs you are not on keep only their icon. |
LoRA start
The All steps button is now Starts at once, or the start you set, such as Starts at 0.75. It opens a bar that stands for the render, from its first step on the left to its last on the right, lit where the LoRA works.
| Preset | Start | What it does |
|---|---|---|
| At once | 1 | The LoRA works on every step. |
| After framing | 0.75 | The model sets the framing and pose alone, then the LoRA joins. |
| Details only | 0.4 | The LoRA joins for the last steps: fine detail and texture. |
Drag the bar or type a number for anything in between.
For agents
| What | Where |
|---|---|
text_projection on set_project |
Picks the projection for a 4B or 8B text encoder. Each take records the projection it used. |
text_encoders and projections in the encoder list |
Small encoders that have a projection, and which projections fit each one. |
Updating
Update the pack and restart ComfyUI so it registers the new node. Projects made with 0.6 need no change. To try the smaller text encoder, add the two files above.
H3 Editor 0.6.1
A small fix to 0.6.0's project zips: opening one can no longer fill the disk.
Fixed
- Opening a project zip cannot fill the disk. Open a project zip wrote the upload to disk with no limit, so a very large or endless upload could fill it. That matters most on a ComfyUI reachable from the internet, such as a cloud GPU. Now the zip must fit twice, once for itself and once for the project it opens into, with 2 GB to spare. A zip that says it is larger is refused before anything is uploaded, an upload that does not say its size stops at that limit, and the answer says how much room the disk has.
| Before | Now |
|---|---|
| Any upload was written to disk until it ended. | An upload stops at (free space minus 2 GB) divided by two, and the part already written is deleted. |
| A too-large zip failed while unpacking, after the upload. | A zip that declares a size that cannot fit is refused at once, with the room there is. |
Updating
Update the pack and restart ComfyUI. Projects and zips made with 0.6.0 need no change.
H3 Editor 0.6.0
H3 Editor 0.6.0 is about working faster in the editor and catching mistakes sooner: project zips to move a project between machines, notes pinned to frames, every check in one list, takes that freeze or go still, pictures for the writing model, right-click menus, and a cleaner side panel.
Highlights
- Project zips. Download a project as one zip and open it on another machine, a cloud GPU or Colab, with every take, only the takes the cut uses, or no renders.
- Notes on frames. Press N on a frame, pin a spot, drag a box or draw, and write what should change. Agents read each note with a picture of its frame.
- Checks in the header. One chip counts what the checks found, and opens a list of it, worst first. Click an entry to go there.
- Frozen and Still marks. The editor now finds where a take stops moving: a held frame, or a stretch where almost nothing moves.
- The writing model sees pictures. Rewrite and Write shots send the cast's pictures, and any picture you add, to the model.
- Right-click menus on clips, VFX clips, sounds, cast members and the viewer.
- A cleaner side panel, with the long explanations behind a ?, and frames and takes shown whole in 9:16 projects.
Project zips
Download as zip and Open a project zip are in the menu beside the project list.
| Level | What the zip holds |
|---|---|
| Everything | Every take, the exports, the pictures and sounds. |
| What the cut uses | The takes the timeline shows, footage, pictures and sounds. Other takes are left out of the project, so nothing points at a missing file. |
| No renders | The shots, the cast with its pictures and voices, sounds, footage and VFX code. The clips open unrendered, with their seeds and settings, ready to render again. |
The dialog shows each level's size before anything is made. Opening a zip makes a new project under its name, or that name with a number, and never overwrites one. Every file in the zip is checked before it is written, and the editor says which LoRAs the project needs that the machine lacks. Uploads of any size work, because the zip is written to disk as it arrives. Thanks to the RunPod user who built this into their template and suggested it.
Checking the cut
| What | What it does |
|---|---|
| Checks chip | Sits in the header and says how many things there are to fix and to check. It opens a list of low cut scores, takes that cut, fade, freeze or go still inside, Soft and Drift clips, VFX clips that pop in or out, and ratings under 70%. Each entry wears its timeline mark's badge and goes to its frame when clicked. |
| Frozen | A red mark where the same frame is held for a few frames or more, such as a take that stalls. |
| Still | An amber mark where almost nothing moves for a second and a half or more. Grain does not count as motion. None of 75 rendered takes we measured was marked. |
| Dissolves | A take that dissolves into another picture must now pass through mixes of the two. A dance or a camera move is no longer reported as a dissolve. |
Notes on frames
Stop on a frame and press N, or Note under the viewer. Pin a spot, drag a box or draw on the picture, and write what should change. With a range marked by I and O, the note covers the range. A note stays on its frame when the clip moves or is trimmed, and says when the clip has a newer take than the one it was written on. Notes show on the ruler and in the Project tab, where they are resolved, reopened or deleted.
Writing with a language model
| What | What it does |
|---|---|
| Cast pictures | Every Rewrite and Write shots request carries the cast's own pictures, each named, so the model sees who it writes about. |
| Your pictures | The dialog takes up to four more: drop, paste, upload, or add the frame the viewer shows. |
| Providers | Works with every provider, ComfyUI's own Generate Text node included. Tested with qwen3.5 9B in ComfyUI and a 27B model behind an OpenAI-compatible server. A model that cannot see writes from the text alone. |
| Model picker | The Model field opens a list in the editor's style: List models fills it, typing narrows it, the arrow keys and Enter pick. You can still type any name. |
Working in the editor
| What | What it does |
|---|---|
| Right-click menus | On a timeline clip, a clip in the list, a VFX clip, a sound, a cast member or the viewer: render, rewrite, extend, bridge, merge, split, show takes, remove, use the frame as a first or last frame, write a note. What cannot be done now stays greyed, with the reason. The menu key or Shift+F10 opens it from the keyboard. |
| Drop on frames | Drop a picture on First frame or Last frame to use it. |
| Side panel | Long explanations hide behind the ? at the end of the panel's tabs and become the tip of their setting. Draft and Final size share a row, and Models and their settings is a link beside Model. |
| 9:16 projects | First and Last frame and the takes show the whole picture, where they used to be cropped to 16:9. |
| Viewer | Hatched outside the picture instead of black, so bars beside a picture of another shape no longer look like part of the video. |
| Version | The header shows which H3 Editor this is, so a screenshot says which release it came from. |
| Live preview | Starts muted, with a button for its sound. |
Fixed
- The viewer went black for a moment at a cut, and at every extension, while the next clip was still seeking to its first frame.
- A project with no notes printed "null" in the Project tab.
- A cast member clicked while a VFX clip was selected did not open.
- The VFX clip's buttons touched the panel's edge, and "frames" wrapped onto its own line after Transition in.
For agents
| Piece | What it does |
|---|---|
add_note, resolve_note, reopen_note, remove_note |
Operations on notes pinned to frames. The project summary lists the notes. |
h3.py notes PROJECT |
Prints each note and saves a picture of its frame with the drawing on it. |
h3.py cuts PROJECT |
Also lists inside: frozen and inside: still, with how long each lasts. |
python cuts.py --still FILE.mp4 |
Lists the frozen and still stretches of any finished file. |
h3.py pack PROJECT --keep cut and h3.py unpack FILE.zip |
Move a project to another machine as one zip. pack --sizes prints each level's size. |
GET /project/{name}/archive and POST /projects/open |
The routes behind them. |
h3.py check |
Now prints the version it talks to. The projects route returns it as release. |
Updating
Update the pack, restart ComfyUI and reload the editor: the project zips need the restart, since ComfyUI adds new routes only when it starts. Projects need no change. The editor starts with the long explanations hidden: press ? at the end of the side panel's tabs to see them.




