Research updates
Every update is dated and condition-bound
Provider claims, project reports, community reports, and pool measurements are kept distinct. This pool has no measured run yet.
Latest
31 October 2026 · Genie 4 access and session limit
Genie 4 can be steered live. It cannot yet run your all-day show.
Google says trusted-tester preview now, later subscriber and partner routes, and only a few minutes of continuous interaction. Public price and viewer delay are unknown.
29 October 2026 · SGF+ route split
SGF+ generated 24-hour videos. The 5090 ComfyUI path is not live.
One B200 browser integration reports 19 generated FPS for 16-FPS playback; a 5090 file job needs 56.6 seconds for 20 seconds. The paper, browser and file answer different questions.
27 October 2026 · SolarWM-H3 browser response
SolarWM-H3 reports 1.15 seconds from keypress to screen—on eight Blackwell GPUs.
The project times a matching browser frame, not only a kernel. P95 response, quality and complete cost remain unknown.
27 October 2026 · EdenLive deployment check
EdenLive beats 24-fps playback on a 5090. Is that the live browser route?
The warm LightX2V MP4 result and the 80-GB-verified WebRTC service need separate tests; moving-object quality remains a gate.
24 October 2026 · Reactor ComfyUI client
ComfyUI can steer hosted FastH3-Omni, but the model still finishes each clip before play.
The node pack needs a Reactor API key. Compare hosted and local routes by receiver delay, reference count and the actual bill.
23 October 2026 · FastH3 Live v2.1.0
One RTX 5090 can beat 24-fps H3 playback on this pinned stack. Can it take viewer direction?
The author reports 10.8 seconds per 15.08-second warm clip; no controlled receiver or full bill was measured.
22 October 2026 · FastH3 Omni release
FastH3 Omni's 12× claim ends at an MP4, not a live receiver.
One warm eight-GB200 job beats its clip duration; a reference-heavy job does not. Local hardware and quality need their own test.
21 October 2026 · TaoMate-H3 vLLM merge
TaoMate-H3 streaming merged. The main-branch stream is video-only.
One-session two/four-H200-class profiles and prompt changes are documented. AAC remains open; a generation-window median is not viewer response or cost.
20 October 2026 · fal Worlds session and price
fal Worlds opens in a browser. The $4.50 figure is compute time, not a delivered hour.
Abot is labeled Commercial use; Lingbot Research only. Both list the same compute-second rate. Viewer response and the full session bill remain unmeasured.
19 October 2026 · H3 benchmark correction
The reported H3 slowdown disappeared. Do not price a route from that CI run.
The four-H100 nightly slowdown returned to normal and was closed; a new API metadata PR has no measured serving gain.
18 October 2026 · Self-Forcing stepwise PR
One A800 streamed its first nine frames in 1.8 seconds. It could not keep up with playback.
An open vLLM-Omni branch has an incremental WebSocket client, but its reported one-A800 chunks fall behind 16-fps playout. Remote viewer delay and cost are unmeasured.
17 October 2026 · H3 first-latent decoder
H3 can decode its first frame in 40 ms. Generation still sets the wait.
A new H3 VAE decoder removes the wait for future latents, but the reported 40 ms begins only after a latent exists; end-to-end viewer delay and cost remain unknown.
16 October 2026 · ComfyStreamerH3 HLS gate
One RTX 5090 beats playback speed. The browser still waits for two clips.
Warm paired-job throughput is not prompt-to-viewer delay. The continuous demo starts after two finished MP4s; managed GPU pricing is a separate route.
15 October 2026 · StreamForce browser control
StreamForce: an unpaced change sat 13 seconds ahead of display in a local test
One H200 can generate ahead of 16-FPS playback, but an unpaced force change sat about 13 seconds ahead of the viewer. Local pacing shortened the lead; long-session quality and remote player delay remain open.
14 October 2026 · Waypoint-1.5 browser route
Waypoint-1.5 plays in a browser on one GPU. The world lasts seconds.
One-GPU browser control works in a project report; local FPS, remote delivery, cold compile and short scene coherence make different claims.
13 October 2026 · StreamRender-H3 browser clocks
StreamRender-H3 has a browser loop. Its 218 ms figure is not player delay.
Eight GB200s, a racing-only renderer, three timing boundaries and no published control-to-screen or full cost.
12 October 2026 · TeleFuser receiver clock
TeleFuser made 60 seconds of media. The client session took 79 seconds.
Four-H100 LingBot-World v2 target compute beat 16 FPS, but one LiveKit replay took 79.25 seconds to complete 59.75 seconds of media.
11 October 2026 · Scope show-readiness
Scope has a browser stream. A paid live show needs more than FPS.
Documented WebRTC input, output and prompt control. One remote-beta event test reports about ten minutes cold readiness; paid-use rights and handover remain open.
10 October 2026 · Hallo-Live model versus client
Hallo-Live reports 20 FPS; its public command saves an MP4
The 20.38-FPS paper throughput is below the public 24-fps file setting and is not a timed viewer session. Default inference peaks around 60 GB GPU memory by maintainer estimate.
9 October 2026 · LiveAvatar route and clock
LiveAvatar: 45 FPS on five H100s, with a longer delivered delay
Five H100s produced the paper result; the authors estimate about three seconds end to end with transport. A one-H200 cold job is an offline clip, not the same service.
8 October 2026 · ShotStream client boundary
ShotStream reports 16 FPS on one H200. Its public example saves a file.
The paper measures model throughput; the listed command reads a prepared CSV and writes a final MP4, not a live receiver session.
7 October 2026 · Veda H3 full-job timing
Veda cut this H3 render to 130 seconds. The video lasts five.
The 7.1× claim times attention only. The full 5070 job improved 2.63× and still cannot feed a live stream.
6 October 2026 · RAVEN H3 memory-envelope check
The 24-GiB H3 preview was tested on an H200, not a 24-GiB card
Its eight-second test used synthetic conditioning and took 301 seconds to sample; the real-prompt test was a separate 39-frame run.
5 October 2026 · Griffin-Lite preview and clock boundary
Griffin's 0.43-second face is not a 0.43-second conversation
Tavus reports a full-duplex preview, but customers cannot use it yet. NVIDIA's event timing uses a different clock from the vendor's generator latency.
4 October 2026 · H3 mask-editing API and client status
H3 mask editing has a server API, but its dedicated ComfyUI node is still open
vLLM-Omni merged request-scoped H3 editing; the ComfyUI PR remained open at the check. This produces a completed clip, not a live edit.
3 October 2026 · Hopper FastH3 complete-clip boundary
Eight H100s finish a 14.375-second H3 clip in 13.506 seconds
Warm, saved MP4 with audio. The separate eight-H200 patch test is not the same benchmark; accepted quality and viewer delivery are unknown.
2 October 2026 · TaoMate-H3 first-chunk boundary
TaoMate-H3 reaches a playable chunk in 17.287 seconds on eight H20s
The ten-second T2AV benchmark is not an end-to-end live session. The official chunked runtime and a separate 5090 LoRA adaptation serve different jobs.
1 October 2026 · NVIDIA Sol-H3 benchmark boundary
Five seconds of H3 in 1.653 seconds. That is not a live stream.
The eight-B300 warm timer excludes final MP4 encoding. NVIDIA lists continuous delivery and direct interaction as future work.
30 September 2026 · Vidu S2 live-editing route
Vidu S2 documents live feed editing at $18 per billed hour before delivery
Its Editing API connects an incoming RTC feed to edited RTC output and optional RTMP relay. Viewer delay and long-session reliability are still unknown.
29 September 2026 · H3 short-clip whole-response check
Four H3 denoising steps still took 11.7 seconds for a 4.4-second clip
vLLM-Omni reports a warmed eight-B300 FastH3 Dense T2VA complete response of 11.7–11.8 seconds. Its 2.36–2.37-second diffusion stage is not the whole wait.
28 September 2026 · vLLM-Omni H3 long-video boundary
H3 can make a longer clip in vLLM-Omni. It is not a live stream.
v0.30.0 adds opt-in Ref2VA continuation. A project test reports 30.667 seconds on eight B300 GPUs; the full wait and viewer-visible delivery are not measured.
27 September 2026 · ComfyUI H3 VAE stage check
ComfyUI reports a faster H3 VAE. The full-render speed is still unknown.
Comfy Org reports a 24.3-to-12.7-second VAE encode/decode A/B on an RTX 5090 at 1344×768, 129 frames. Denoising and delivery were outside the timing.
26 September 2026 · vLLM-Omni H3 batching boundary
H3 can schedule between denoise steps. Its four-request test still took longer.
vLLM-Omni’s stated two-H100 test found four simultaneous H3 jobs slower with four-way co-batching than ordinary request mode. Its current two-RTX-5090 DLO launch also selects the FL2VA task path. Neither result is a live route.
25 September 2026 · vLLM-Omni client-scope check
The H3 server can take first and last frames. The current vLLM-Omni ComfyUI client cannot send both.
The serving recipe documents ordered two-keyframe conditioning. The current H3 API/ComfyUI RFC says the current client accepts one FL2VA frame and constrains references. Server support is not a workflow result.
24 September 2026 · vLLM-Omni two-RTX-5090 capacity check
Two RTX 5090s can run full H3. The documented job still takes 8:38 for five seconds of video.
vLLM-Omni says one 50-step T2VA job returned 124 1344×768 / 24 fps frames in 8 minutes 38 seconds on two RTX 5090s. Its DLO route needs at least 200 GiB host RAM. This is offline capacity, not a live route.
23 September 2026 · vLLM-Omni FastH3 timing check
Eight B300s can finish a FastH3 clip before playback ends. That still is not a live stream.
vLLM-Omni says a warmed eight-B300 FastH3 T2VA profile returned 10.125 seconds of complete MP4 in 8.678–8.710 seconds. The report does not measure streaming, first frame, cold start, queue, or viewer-visible response; its raw bundle is pending.
22 September 2026 · Dreamverse FastH3 profile check
Dreamverse gives FastH3 a four-GPU chain. It does not give it a live-capacity result.
FastVideo documents 124-frame 768×1344 Preview v1 segments, last-frame carry, a four-GPU default, and a cold readiness wait that can take tens of minutes. It is a continuity mechanism to inspect, not a stream benchmark.
21 September 2026 · FastH3 V2 native-canvas check
FastH3 V2 has a speed boundary before it has a speed claim.
One detailed Radeon 8060S community report puts a 768×1344, 243-frame V2 run at 33 minutes 9 seconds. The author says the sparse path does not help below a stated token threshold; this is local rendering, not live video.
20 September 2026 · FastH3 V2 scope recheck
The FastH3 V2 transition claim is still a source conflict.
FastVideo’s model card says first/last-frame conditioning was not distilled. ComfyUI’s current template says it accepts those frames. Do not call it a continuity route before a pinned render proves the anchors hold.
19 September 2026 · RTX 5090 FastH3 timing check
A 5090 FastH3 run is a faster review loop, not live video.
A platform-published FastH3 Preview v1 test returned 15.08 seconds of image-to-video in 55 seconds at 480×640 and 126 seconds at 768×1024. Its hot timing excludes queue and model loading.
18 September 2026 · FastH3 Comfy scope check
FastH3 reached ComfyUI. Its first/last-frame claim is already disputed.
FastVideo documents a fixed eight-forward V2 recipe as T2AV-only; a current official ComfyUI template says it also accepts first and last frames. Do not promise a transition before a pinned run settles it.
17 September 2026 · Remote H3 serving reality check
An OpenAI video endpoint does not make H3 a live-video server.
vLLM-Omni documents a separate H3 service that ComfyUI can call. Its stated full-H3 four-B300 row returns about 8.7 seconds of output in about 87 seconds.
16 September 2026 · Collective control-loop check
YoLive has a real-time claim. The useful question is who gets to choose the next scene.
Yoroll says group proposals and votes guide the next scene, with 10 seconds of 768p, 24 fps video generated in four seconds on eight B200 GPUs. The missing measurement is vote-to-screen time.
15 September 2026 · Consumer-channel reality check
A 5090 can reportedly keep FastH3 on air. That is not the same as a live 24-fps show.
A community source reports 22.1 fps at 448×448 on one RTX 5090. Its continuous channel relies on retimed playback, prequeued work, and distinct scenes.
14 September 2026 · Capacity correction
Four B200s do not clear FastH3’s documented 15-second deadline
Reactor gives separate four- and eight-B200 timings. The latter is faster than the cited clip duration; neither is an independent live-service result.
13 September 2026 · Autoplay boundary
FastH3 can autoplay. Its default queue does not.
Reactor documents optional hands-free playout, but a completed clip otherwise clears to black. The queue policy is part of the product.
How to read a record
- Provider says proves a published capability, price, or product term.
- Project says records a project’s stated setup or measurement.
- Community report records a user’s stated experience and is a lead for replication.
- Measured here will appear only with a pool test, method, inputs, and result.