Hy-Nomas: a 4090 fit claim, an eight-B200 browser test
Project and community reports: the low-memory RTX 4090 preset has no published viewer rate. The eight-B200 Reactor route generated 18.53 FPS over two minutes; a separate 4090 ComfyUI job took 85 seconds for 10 seconds of 1280×704 video. Decoder-weight rights for paid use need clarification. Compare the routes.
Micro-World 2: 42.49 decoded FPS on one MI350X, not a priced viewer session
AMD reports: its 352×640 action-controlled model reaches 42.49 decoded FPS with a lightweight decoder or 32.12 with the original Wan VAE. The faster decoder reduces visual quality; no browser response, accepted long session or full bill is measured. Read the route and decoder check.
EdenLive on a 5090: 25.17 FPS in a file, not yet a browser result
Project reports: LightX2V’s warm 717-frame, 1280×704 proxy-video job runs at 25.17 inference FPS on one RTX 5090. It saves an MP4. EdenLive’s WebRTC service guide lists an 80-GB GPU in its verified environment; action-to-screen delay and a full bill remain unmeasured. Check the two deployment paths.
FastH3 Omni: faster finished clips on eight GB200s, not proven live
Project says: one 15-second, single-image job finished as an MP4 in 14.14 seconds; a 12-reference job took 59.77 seconds on the same eight-GB200 benchmark. Local 5090/ComfyUI setups use different profiles. No receiver delay or full bill is measured. Compare the routes and clocks.
TaoMate-H3: merged video, not yet synchronized AV
Project says: vLLM-Omni merged a one-session WebSocket route on 19 October. The two-H200-class FP8 test reports a 4.74-second median per 4.958-second window, but that is not viewer response. The merged MP4 carries video only; AAC work remains open. Check the route and limits.
Latest check: ComfyStreamerH3 on one RTX 5090
Project reports: two 15.08-second 448×256 clips in a 23.51-second warm paired job. Its HLS browser waits for two completed clips and queues a new prompt until the current clip ends. The viewer response is unmeasured; the managed example uses a different GPU and price. Read the two-clock check.
Scope: a browser route with a paid-show gate
Project says: its WebRTC guide sends source video, returns generated video and accepts live prompt changes. Community report: one remote-beta H100 event test took about ten minutes to cold-ready; LongLive T2V and StreamDiffusionV2 V2V timings are not interchangeable. The repository alpha license is non-commercial, and event delay, one-hour rollover and full cost remain open. Read the show-readiness check.
Hallo-Live: test a spoken-scene model, not a turnkey client
The two-H200 paper reports 20.38 generated FPS. Its public inference command saves a 24-fps file, above the paper throughput, and a maintainer estimates about 60 GB peak GPU memory for default inference. Treat 24 GB as a theoretical target, not a setup recommendation. Read the pre-spend test.
LiveAvatar: choose clip or conversation
Authors report: five H100s, 45.2 generated FPS and 1.21 seconds from audio arrival to first visual output at 720×400. They estimate roughly three seconds with transport. The released one-GPU command is offline; an independent H200 cold finished clip took 181.44 seconds. Do not price a live conversation from that clip. Read the route and clock check.
ShotStream is a research route, not a packaged live client
Paper reports: 15.95 generated FPS at 832×480 on one H200 for multi-shot stories. The public example reads prepared prompts from a CSV and writes a final MP4. Use the reference checkpoint for a bounded offline study; do not choose it as a viewer-ready live setup without a prompt-change and receiver test. Read the boundary.
Decision guide · checked 7 October 2026
Choose the control and deployment contracts, then test the claim
A model name does not tell you whether you are buying a group-choice loop, an interactive session, a remote GPU service, a clip queue, an unattended local channel, or a short local render.
New watchlist: Tavus Griffin-Lite
It is a selected-tester research preview for full-duplex face-to-face video. Tavus reports 0.43-second audio-to-video generator latency on H100s, not a customer session. No Griffin-specific customer price or public route is documented. Read the access and timing check.
Veda: test a faster offline H3 clip
The project reports a complete 5.17-second H3 T2VA clip in 130.1 seconds on one RTX 5070 with Veda, versus 342.1 seconds on its default ComfyUI path. This is a viable acceleration experiment, not a consumer live-video route. Match task, version, output and quality before switching. Read the timing card.
Choose the job before the route
| Need | Route to inspect | Evidence status | Constraint that changes the answer |
|---|---|---|---|
| Warm H3 clips on Hopper | Hopper FastH3 patched route | Project says: eight H100 80 GB GPUs saved 14.375 seconds of 1344×768/24 fps audio-video in 13.506 seconds. | Four-forward sparse checkpoint. Load, quality acceptance, viewer delivery, retries and cost unmeasured. A separate matched A/B used eight H200s. |
| A chunked H3 T2AV experiment on datacenter GPUs | TaoMate-H3 runtime | Project says: first playable benchmark video in 17.287 s for a ten-second 480×864 job on eight H20 96 GB GPUs. | Excludes internal audio preparation; original benchmark command saved an MP4. A separate vLLM video WebSocket route has since merged; neither has a pool receiver or cost result. The three-step LoRA alone is not the streaming runtime. |
| Fast offline H3 clips on a resident multi-GPU system | NVIDIA Sol-H3 | Project says: warm eight-B300 1344×768 T2VA clips beat nominal playback duration through VAE decode. | Four-forward adapter; final encoding, startup, viewer delivery and continuous control unmeasured. Validate quality and bill before sizing a show. |
| Change the look of an incoming live feed | Vidu S2-Editing | Provider says: RTC input and edited RTC output, with optional RTMP relay and mid-stream reference changes. | One credit per billed live second; standard $0.005 credit implies $18 per hour before delivery. Control-to-viewer delay and sustained behavior are unmeasured. |
| A longer completed H3 clip, not a live response | vLLM-Omni v0.30.0 Ref2VA continuation | Project says: an opt-in latent-tail route produced one 30.667-second original-H3 output on eight B300 GPUs at 1344×768, 24 fps, 50 steps. | Full wall time, first playable frame, joins, cost, and repeated-run reliability remain unknown. The 300-second request limit is not a tested output length. |
| A remote H3 service for a client or team | vLLM-Omni H3 server | Project says: the current two-RTX-5090 DLO launch selects the FL2VA path, which covers T2VA and FL2VA. Its stated two-H100 test found four simultaneous H3 jobs slower with four-way co-batching than ordinary request mode. One T2VA validation completed on two RTX 5090s in 8:38 for 5.17 seconds of output; a separate warmed FastH3 T2VA profile returned a complete 10.125-second MP4 before playback ended on eight B300 GPUs. | Task, exact adapter artifact and schedule, server and client revision, arrival pattern, GPU and host memory, cold/warm time, first frame, client compatibility, queue, delivery, and cost. The two-5090 route recommends a 384 GiB host; neither row is a live screen result. |
| A shared audience should select the next story beat | Yoroll H3 Superfast / YoLive | Provider says: proposals and votes guide the next scene; 10 seconds at 768p/24 fps with native audio generate in four seconds on eight B200 GPUs. | Vote window, moderation, input-to-screen delay, public uptime, cost, and what survives into the next scene. |
| Viewer input should alter a held stream | fal H3 Max Director | Provider says: persistent WebRTC tracks, carried context, and mid-session text directions. | Public session limit, current $0.08/second billing, reconnect, and moderation design. |
| A frame-carried FastH3 creative loop | FastVideo Dreamverse FastH3 profile | Project says: four visible GPUs by default, 124-frame 768×1344 audio-video segments, and last-frame conditioning into the next segment. | Cold readiness, warm segment time, conditioning fidelity, ready-buffer margin, recovery, cost, and the visible response to a direction. |
| A chat-shaped feed of short scenes | FastH3 queue and playout | Project says: separate build and play queues can feed paced output; its cited 15-second build is 15.5 s on four B200s and 12.9 s on eight. | GPU profile, warm-up, ready-buffer depth, autoplay/filler policy, hard cuts, and what viewers see after a miss. |
| An unattended local AI-TV channel | One-RTX-5090 FastH3 community route | Community project says: its named dense FastH3 workflow sustains 22.1 fps at 448×448. | Retimed motion, source variant, attention path, prequeue, scene independence, host memory, image quality, and repeatability. |
| A local FastH3 clip or possibly a frame-anchored transition | FastH3 8-Step V2 ComfyUI template | Project and ComfyUI say: a new eight-forward package and template exist. | Scope conflict: FastVideo says T2AV only; the template says T2VA and FL2VA, not Ref2VA. No local timing, quality, or hardware result was found. |
| One controlled local shot | MiniMax H3 through a local framework such as ComfyUI | Provider and community reports say: local base clips and an optimized dynamic-VRAM path exist; one 5090 report gives a condition-rich offline result. | Task mode, system RAM, model pack, resolution, steps, software version, and wait tolerance. |
Pick remote serving when
You need the creator UI and the H3 machine to be different systems, or you want several clients to reuse one service. vLLM-Omni’s current H3 recipe records an OpenAI-compatible route. Its current H3 API/ComfyUI RFC says the server and current client have different input limits: check that the client can send the needed keyframes or references. The recipe documents one two-RTX-5090, 50-step T2VA job at 1344×768: 124 frames at 24 fps, about 5.17 seconds of output, returned in 8 minutes 38 seconds. The current two-GPU launch selects the FL2VA path, which serves T2VA and FL2VA, and recommends a 384 GiB host. That is an offline capacity lead. A separate eight-B300 FastH3 T2VA report says a complete 10.125-second MP4 returns in 8.678–8.710 seconds after warmup. Keep the two rows separate.
Pick an offline two-5090 H3 test when
You need one controlled T2VA or FL2VA render and can supply server-class host memory, storage, and a long returned-artifact wait. The published DLO launch selects FL2VA; the documented record is still T2VA only. Validate a first/last-frame request and a second queued request before renting or buying for a continuity product.
Do not sell a queue with a scheduler flag
Project says: its stated two-H100 H3 test submitted four jobs at once. Four-way co-batching took longer than ordinary request mode. The source leaves staggered arrivals unmeasured. State the arrival pattern, job shape, concurrency, first playable media, and fallback before you call the route responsive.
Pick a group-choice loop when
You want participation to create anticipation for the next scene instead of a viewer controlling the current one. Yoroll’s launch announcement says YoLive uses proposals and votes. The stated generation time is fast, but do not choose the route until you know the selection rule, moderation delay, response time, delivery behavior, and cost.
Pick Dreamverse when
You need a published FastH3 application design that tries to carry a visual boundary forward rather than merely play distinct ready clips. The current Dreamverse README calls for four visible GPUs by default. Treat it as a four-GPU evaluation route: time the cold readiness state, then inspect a warm handoff before you write a continuity promise.
Pick Director when
You need the product shape of an open session. The current provider page says it streams through WebRTC and carries context into later directions. Plan for a finite session from the first sketch, and price $4.80 for even the stated 60-second minimum.
RAVEN on a 24-GiB card: not yet demonstrated
The maintainer's 192-frame record ran on an H200 with 24.100 GiB left free, not on a physical 24-GiB GPU. It used synthetic conditioning. A separate 39-frame real-prompt test ran without the memory limit. The maintainer advises 160 GiB available host RAM and verified ComfyUI 0.33.0. Do not choose hardware or promise live playout from the 24-GiB headline. Read the claim check.
Baseline card before you spend a weekend
- Write who can propose, who selects, what changes, and whether the active scene can change.
- State client host, inference host, task, adapter load point, server revision, GPU, host memory, and any carried conditioning frame.
- Record cold, warm, accepted-input, ready, HTTP response, delivered-start, and visible-change timestamps.
- For a channel or queue, record ready clips, playback speed, filler, retries, and black or silent fallback.
- Label the result as provider documentation, project report, community report, or a measured test. Do not blend them.