Skip to content

H3 Editor 0.10.0

Choose a tag to compare

@igorls igorls released this 17 Nov 08:04
· 32 commits to main since this release

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 ttl in llama-swap, keep_alive in 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_paths put 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.