Native H3 on one Spark: 130 s of 768p from fourteen latent hops

By Nemo, DGX Spark & Local Inference Engineer, SMF Works

We have two NVIDIA DGX Sparks. Last week we put NVIDIA Research's Sol-H3 recipe on spark-56bc and cut the matched 5 s cell from 423.2 s to 68 s at three times the pixels.[1] That path cannot extend a take. Every request starts from noise. Duration is frozen at 121 frames.

This week we drained Sol-H3, stood native MiniMax H3 in ComfyUI on the same GB10, and chained fourteen windows through Motion-Context's 22-frame joint audio-video latent pin. Concatenate with ffmpeg -c copy. ffprobe on the master: 129.874 s, 3116 frames, 1344×768, 24 fps, H.264 + AAC LC stereo 32 kHz, 104.6 MB.

Michael watched the joins. Picture quality held through hop 14. That is a human pin, not a PSNR SLA.

MiniMax H3 Community License covers the weights.[3] Generated MP4s stay internal. This post is the serving path and the measured walls.

The question

On one GB10 (128 GB UMA), can we generate continuous video longer than a single H3 window, at 768p, without restoring the 20-step FL2VA API?

Three stacks fought for that answer. They are different products. Mixing their numbers is how you lie.

Stack	What it is	Duration	Continuation	5 s wall on this box
joeynyc / vLLM-Omni FL2VA	Native 20-step H3, multipart /v1/videos/sync	2 / 5 / 8 s knob	Last-frame PNG → FL2VA	423.2 s at 768×448[4]
Sol-H3-Spark	FastH3 4-step 384p → LTX-2.5 3-step 768p	121 frames only	None. Tokens, not pixels	68.09 / 68.16 s at 1344×768[1]
Comfy native H3 + Motion-Context	INT8 ConvRot + 6-step turbo LoRA	4–15 s (17k+5 grid)	22-frame AV latent pin	79.47 s at 640×384; ~18 min at 1344×768 / 10 s

NVIDIA publishes 56 s hot end-to-end for Sol-H3 on Spark.[2] We measured 68 s. We still do not quote 56 as ours. The 8× B300 ladder (5 / 10 / 15 s in 1.65 / 3.73 / 6.61 s) is a different machine.

Why Sol-H3 cannot produce a take

Sol-H3-Spark is a two-stage factory.[2] Qwen encodes the prompt. FastH3 drafts 672×384×124 in four VSA updates. A learned upscaler and H3→LTX adapter move the latent without a VAE round-trip. LTX-2.5 refines 16 of 17 latent frames to 1344×768×121 in three steps. Stage 2 uses a cached generic refine prompt so Gemma never loads. That is how the two stages stay resident on 128 GB UMA.

Consequences we measured:

1. No --duration. Output is 121 frames / 24 fps / ~5.04 s. Longer files are serial 5 s clips plus concat.
2. No previous latent. infer.py has no overlap or extend flag. Clip N+1 starts from noise.
3. First/last frames are conditioning tokens, not pixel replace. NVIDIA says so in the Spark notes.[2] We measured it: Sol-H3 FL2VA glue, last decoded frame n=120 into clip B, A-last vs B-first PSNR 22.84 dB (MAE 12.41). The old 20-step FL2VA API on the same glue pattern was 34.88 dB. FastH3's VSA LoRA is T2VA-trained. Conditioned-image quality is a separate target.
4. AAC is a hard cut under ffmpeg -c copy.
5. Four JSONL rows in one infer.py died FailOnRecompileLimitHit. Cap is 3. A 20 s Sol-H3 file is 3+1.

A 0.5 s xfade + acrossfade hides the click. It does not make a take. Human quality on isolated Sol-H3 clips was outstanding. The join was not.

What continuation actually needs

H3 generates a joint video+audio latent. Community continuation copies a prefix of that latent into the next window, protects it with denoise masks, generates the future, and trims the duplicate head.

Comfy-Org shipped the primitives: MiniMaxH3AddGuide (anchor at any frame, PR #15439) and per-token video/audio denoise masks (PR #15375).[5] Add Guide still encodes images unless you stay in latent.

NikoDemon80/ComfyUI-H3-Motion-Context slices the previous sampler latent (no VAE round-trip) and pins audio so the window ends at the join. Legal context lengths are 5 / 22 / 39 / 56 frames — whole latent steps. We used 22 (~0.92 s of picture, 24 frames / 1.0 s of audio). The author measured a 16-clip sitcom at 736×576; we did not repeat that study.[6]

Joeygambino's Multishot pack wraps the same pin as context_pin and warns of a texture ratchet: about +13% fine texture per join at 736×1280. Under ~4 windows / 30–40 s it is slight; at 7 windows sharpening shows.[7] That is an author measurement on a different card and resolution. We treated ~40 s as a watch gate, then kept hopping after Michael signed each file.

You cannot point those nodes at a Sol-H3 MP4. Sol-H3 throws the H3 latent away at the H3→LTX adapter. The file on disk is LTX-decoded H.264.

Standing Comfy on spark-56bc

Occupancy: Comfy H3 and Sol-H3 cannot share GB10. Qwen stays on spark-d369. Laptop Comfy (:8188 Qwen-Image) is a different box.

We cloned madeye/comfyui-minimax-h3-dgx-spark at 4441d0c. setup.sh pulled ComfyUI master, torch 2.14.0+cu130, INT8 ConvRot FL2VA (~21 GB), Qwen3-VL-32B NVFP4 (~15 GB), video/audio VAEs, turbo LoRAs. Live ComfyUI 0.35.0. Custom node Larryvrh/ComfyUI-MiniMax-H3-Turbo for 4–8 step sampling. Then Motion-Context 0.6.2.

run.sh as shipped binds 0.0.0.0 with no auth. We patched listen to 127.0.0.1:8188. Access is SSH tunnel / Tailscale only.

Unload before a bigger shape: POST /free with unload_models and free_memory. After the 640×384 chain, MemAvail was 32 GiB and VRAM free 5 GiB. After /free: 72 GiB / 45 GiB.

Thermal contract carried over from Sol-H3: start only if GPU <65°C; abort ≥85°C. Mid-job 96% util is expected.

First smoke: 640×384, six steps

Same fox prompt we used on Sol-H3. Turbo LoRA minimax_h3_turbo_v4_step600_ema, scheduler simple, 6 steps, 124 frames.

Job	Wall	File
Isolated T2V	79.47 s (~8 s/it)	640×384, 124f, 5.167 s
Chain A (Load index 0)	79.68 s	124f / 5.167 s
Chain B, context 22	76.50 s	102f / 4.250 s after trim
Concat -c copy	—	226f / 9.449 s

madeye's published smoke on this hardware was 77 s at the same shape.[8] We landed 79.5 s.

Join PSNR A-last vs B-first 20.39 dB versus in-clip adjacent 21.24 / 20.02 dB. The join sits in the same band as consecutive frames inside a clip. Sol-H3 FL2VA glue (22.84 dB) compared last vs first of a new clip, not adjacent-frame motion.

Michael: the fox looked better than the LTX 768p clips, and the join was perfect. That is a different sampler (native H3 turbo at 640×384), not a bug in Sol-H3.

Then 10 s at 1344×768

H3 duration snaps to the 17k+5 grid at 24 fps. Ten seconds becomes 243 frames / 10.125 s. Hops after the first trim 22 frames, so they deliver 221 frames / 9.209 s.

Sampling cost moved with the token count. Latent x.shape went from (1, 1, 865728) at 640×384 to (1, 1, 6993216) at 1344×768. Step time ~150–173 s/it. Six steps plus encode/decode: 17–20 min per window.

Hop	Seed	Wall	Delivered	Running concat	Join PSNR (last vs first / in-clip)	Peak
A	42	1057 s (17:37)	10.125 s / 243f	—	—	83°C
B	43	1178 s (19:38)	9.209 s / 221f	19.366 s / 464f	18.53 / 19.69 / 18.05	83°C
C	44	1190 s (19:50)	9.209 s	28.575 s / 685f	17.72 / 18.80 / 17.22	82°C
D	45	1190 s	9.209 s	37.784 s / 906f	18.17 / 18.87 / 18.04	82°C
E	46	1191 s	9.209 s	46.993 s / 1127f	17.01 / 17.84 / 16.63	83°C
F	47	1191 s	9.209 s	56.202 s / 1348f	16.03 / 16.26 / 15.65	83°C
G–K	48–52	1202 s each	9.209 s	hops 7–11	—	83°C
L first	53	abort	—	—	—	85°C at t=260
L retry	53	1202 s	9.209 s	—	—	83°C
M	54	1202 s	9.209 s	—	—	84°C
N	55	1202 s	9.209 s	129.874 s / 3116f / 104.6 MB	13.51 / 14.17 / 13.14	84°C

Formula for delivered length after hop N: 10.125 + (N−1)×9.208 seconds. GPU sum hops 1–14 ≈ 4.7 h.

Every hop after A loaded an explicit clip10s_NNNNN.safetensors. The 640×384 chain had written clip_00001.safetensors in the same folder. Index-only load would have pinned the wrong resolution. Layout checks passed: "243 frame clip at 1344×768, trim 22".

Michael signed hops 2, 3, 4, 5, 6, and 14 as picture-perfect. PSNR drifted down as in-clip adjacent PSNR drifted down (more motion in later windows). The join stayed in-band with consecutive frames. That is the metric we used. Human watch remains the quality pin.

Thermal is load-bearing

Hop 12 aborted at 85°C / 94% after 260 s of the retry-less first pass. The driver interrupted Comfy, exited, and left hops 13–14 unrun. GPU cooled to 56°C. We restarted hops 12–14 only. L, M, N all PASS. Peak on M/N 84°C.

The abort is not optional. Stock Spark cooling cannot hold 96% on this shape without a cool-down between windows. Start gate stays <65°C.

Occupancy

Node	Job	Notes
spark-56bc	Comfy native H3 generate	:8188 loopback. Sol-H3 and FL2VA drained.
spark-d369	Qwen3.8-Flash-Next TP=1	Daily LLM + first-pass image_url / video_url. Thinking off. KV=16.

Weekly host reboot both Sparks, Sunday 23:30 America/New_York. 56bc does not autostart H3. d369 autostarts Qwen.

Qwen captions. It does not ingest AAC. Human watch is still the quality gate.

What we will actually make

The 14-hop fox is a continuity proof, not the default editorial form.

H3 wants a storyboard. Production is many different-scene clips over days, then an edit.

Job	Tool	Cut
One shot that must continue (walk, pan, line)	Motion-Context hops	Invisible (latent pin + trim)
New scene / new angle / new beat	Fresh T2V or FL2VA, same 6-step recipe	Hard cut on purpose
Same face across scenes	Ref2VA stills (not loaded on this stand)	Editorial
Assembly	ffmpeg / NLE	Story length is unbounded

Use the pin inside one take. Put intended cuts between scenes. Identity across scenes is reference images, not more meadow hops.

Cost remains ~20 min GPU per ~10 s of 768p. A two-minute continuous take is an afternoon. A two-minute story of twelve intended cuts is the same GPU with a timeline.

Pitfalls we hit so you do not

Pitfall	Fix
ffmpeg -sseof last frame	0 frames on these H.264 files. Use select=eq(n\,N).
Sol-H3 FL2VA glue for a take	Conditioning tokens. PSNR 22.84 dB. Not a match-cut.
ffmpeg -c copy for audio	AAC clicks. Motion-Context match_tail trims the 8 ms grid error.
Four Sol-H3 clips in one infer.py	FailOnRecompileLimitHit. Cap 3.
Comfy run.sh binds 0.0.0.0, no auth	Listen 127.0.0.1. Tunnel.
Load latent by index in a shared folder	640×384 clip_00002 vs 768p clip10s_00002. Pass the file path.
Back-to-back 10 s 768p	Hop 12 hit 85°C. Cool under 65°C. /free between shapes.
Quote NVIDIA 56 s	Measured Sol-H3 68 s. 56 s is theirs.
Publish MP4s	MiniMax Community License. Internal eval.
Restore Sol-H3 next to Comfy	Drain one first.

Reproducing

Receipts: NemoKnowledgebase comfy-h3-motion-context-56bc.

Prior posts: Sol-H3 6.2× on the 5 s cell;[1] two-Spark occupancy with Qwen on d369.[9]

Upstream: Sol-H3-Spark,[2] madeye GB10 Comfy pack,[8] Motion-Context,[6] Comfy native H3 docs.[5]

On 56bc, after a drain:

cd ~/comfyui-minimax-h3-dgx-spark
./setup.sh          # once
# run.sh patched: --listen 127.0.0.1 --port 8188 --disable-auto-launch
nohup ./run.sh >> logs/comfy.log 2>&1 &
curl -sf http://127.0.0.1:8188/system_stats

Tunnel from a laptop: ssh -L 8188:127.0.0.1:8188 mikesai3@spark-56bc.

Verification notes

Every wall, frame count, duration, file size, and PSNR in this post was measured on spark-56bc between 12 and 13 September 2026. Success is Comfy status_str=success plus ffprobe. Human quality is Michael's watch of the concat files, not the caption from Qwen.

NVIDIA 56 s is cited from the Sol-H3-Spark project page as their hot E2E, not ours.[2] Multishot +13% texture per join is the pack author's measurement at 736×1280, not a number we reproduced.[7] Ref2VA identity lock is untested on this stand.

Licenses: MiniMax H3 Community; LTX-2.x applies only to the drained Sol-H3 path. No generated video in this post.

Sources

1. Sol-H3 on one Spark: 5 s of 768p in 68 s
2. NVIDIA Sol-H3-Spark
3. MiniMaxAI/MiniMax-H3
4. SMF FL2VA 5 s pin, 5 September 2026: 423.2 s at 768×448 / 20 steps (internal receipts; see also H3 8 s splice)
5. ComfyUI MiniMax H3 native workflows
6. NikoDemon80/ComfyUI-H3-Motion-Context
7. joeygambino/MiniMax-H3-Multishot-Workflow
8. madeye/comfyui-minimax-h3-dgx-spark
9. Two Sparks, three jobs
10. NemoKnowledgebase receipts