+ 
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ 4.5 s requested → rounded up to 124 frames = 5.167 s at 24 fps
+ (legal shapes are 5 + 17n, 22…362 frames). The two flags are mutually exclusive.
+
+
+
+
+
+ 16:91:1
+ 9:164:33:4
+
+
+ 256²512²
+ 768²1344×768768×1344
+
+ 512 × 512 = 0.26 MP of the 1.03 MP limit (768 × 1344).
+ Both sides must be multiples of 32 and at least 32.
+
+
+
+
+
+ DraftBalanced
+ Reference
+
+ Balanced =
+ --steps 20 --layers 45 --reuse 2 --token-reduction.
+ Estimated 6 min on this GPU. Change any of it in Advanced.
+
+
+
+
+ Decode one frame after every denoising step — adds a
+ preview VAE load
+ phase and one decode per step.
+
+
+
+
+ drop an image
choose from library
+ choose from library
+
+
+ drop an image
choose from library
+ choose from library
Frame anchors select the FL2VA path and
+ cannot be combined with ordered references. Clear the 5 references to enable them.
+
+
+
+
+
+
+ random
+ Command this will run
+./h3 -d $H3_MODEL_DIR -p "A red fox walks through fresh snow…" --width 512 --height 512 --frames 124 --steps 20 --layers 45 --reuse 2 --token-reduction --seed 42 --preview-dir data/jobs/13/preview -o data/jobs/13/out.mp4
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ Output canvas
+
+
+
+
+ Multiples of 32, at least 32, product ≤ 768 × 1344 = 1 032 192 px.
+
+ Lower internal canvas --render-width --render-height
+
+
+
+ Set both or neither, same aspect ratio as the output, multiples of 32,
+ never larger than the output. Model and VAE run at this size, then upscale.
+
+
+
+
+ Assigned by the server. Empty disables MP4 encoding (
+ -o '').
+ Write frames as PPM --frames-dir
+
+ Final frames only; denoising previews are not written here.
+ Profile phases --profile
+
+ Per-phase wall time, peak tensor storage, dispatch counts.
Duration
+
+
+
+
+
+
+ Mutually exclusive. 4.5 s → 124 frames = 5.167 s. Requests round up to
+ 5 + 17n; below 22 frames generation is refused (one 22-frame decoder chunk minimum),
+ above 362 frames it is out of range.
+ Sampler
+
+
+
+
+ 2 … 1000
+
+ 35 … 50 — 50 exact, 45 fast, 40 aggressive
+
+
+
+ 1 close, 2 fast, 3 aggressive
+
+ 1 exact, 4 fast, 6 aggressive
Mutually exclusive: core reuse and denoiser reuse cannot both exceed 1.
+
+ Token reduction --token-reduction
+
+ Pairs horizontal video tokens in middle DiT blocks. Keep off at 256 square,
+ and do not combine with both
--layers 40 and --reuse 3.Memory and backend
+
+ SSD streaming --ssd-streaming
+
+ Exact low-memory mode: GB10 DiT peak 27.06 GB → 1.63 GB, load+denoise +37.6 %.
+ int8 row FC2 --use-int8-row-fc2
+
+ Metal/M5 only — a measured no-op on this CUDA backend. Also incompatible with
+ SSD streaming and with the BF16 MLP.
+ Reference RoPE --use-reference-rope
+
+ Restores the released 256×256 spatial grid for parity checks.
Parity and debug flags 10
+
Force close-reference implementations. Slower by
+ design; useful only to reproduce reference numerics.
+
+
+ --use-slower-bf16-mlp
+ --use-slower-bf16-qkv
+ --use-slower-bf16-attention-output
+ --use-slower-row-major-attention-output
+ --use-slower-unfused-int8-inputs
+ --use-slower-unfused-qkv-rope
+ --use-slower-scalar-qkv-rms
+ --use-slower-uncached-int8-scales
+ --use-slower-dynamic-fc1-k
+ --use-slower-grouped-quantizer
+ Post-processing
+
+ Face replacement
+ faceswap
+ Unavailable: no model and no runtime installed. Set
H3_FACESWAP_CMD
+ to an executable to enable it. This repository ships neither models nor download URLs;
+ known model licences are non-commercial/research only.Not exposed by this UI 5
+
-
+
- --model-dir and --info — server configuration, from
H3_MODEL_DIR.
+ - --show and --zoom — Kitty/Ghostty terminal graphics; the browser uses --preview-dir instead. +
- --help — CLI only. +
+
+
+
+
+ matchmax
+
+
+
+
+ Command-line order is preserved and matters.
+ Ordered references select the Ref2VA checkpoint.
+
+
+ imagefox.png
+ 720×720 · --ref-image↑ ↓ ✕
+ videoclip.mp4
+ keeps audio · --ref-video↑ ↓ ✕
+ videosilent.mp4
+ drops audio · --ref-silent-video↑ ↓ ✕
+ vid+audscene.mp4 + music.wav
+ --ref-video-audio↑ ↓ ✕
+ audiomusic.wav
+ 6.2 s · --ref-audio↑ ↓ ✕
+
+ total 5/12images 1/9
+ videos 3/3audio inputs 3/3
+ audio duration 10.4/15 s
+
+
+ + image+ video
+ + silent video+ video + audio
+ + audiochoose from library
+
+ -
+
- At most 12 references: 9 images, 3 videos, 3 audio inputs. +
- Audio needs at least one image or video reference alongside it. +
- Standalone audio must last at least 2 s at 32 kHz; total reference audio ≤ 15 s. +
- A video soundtrack is truncated to the output duration and needs ≥ 2 s, so request at least 56 output frames. +
- References cannot be combined with --first-frame / --last-frame. +
Blocked: videos 3/3 and audio inputs 3/3 are full — remove one before adding another.
+
+
+
+
+
+
+
+
+
+
+Queue
+
+
+ running
+ #13 fox in snowcancel
+ live preview
denoise step 7 of 20
+
+
+ denoise step 7 of 20
+
+ queued
+ #14 surfer inside a waveremove
+
+
+
+ failed
+ #12 portrait 1400×800dismiss
+ h3: canvas exceeds the released 768*1344 pixel limit
+
+
+
+ done
+ #11 red cube on whitedelete
+
+
+ One job runs at a time: a single GPU, 27 GB DiT peak.
+ Library
+
+
+
+
+
+
+
+
+ fox.png · 720×720
portrait.jpg · 1024×1536
clip.mp4 · 4.2 s
music.wav · 6.2 s
Uploaded assets stay selectable for later jobs as
+ first/last frame or as ordered references.
+ Gallery
+
+
+
+
+
+
+
+
+ #11 cube · 00:41
#09 surfer · 06:12
#08 hummingbird · 05:48
#07 fox 256² · 01:03
' opacity='0.85'/%3E%3C/svg%3E)
+ 
+ 
+ 
+ 
