One RTX 5090: 24-fps H3 playout, not measured viewer control

Community project reports: FastH3 Live v2.1.0 makes each 15.08-second 448×448 video-and-audio clip in 10.8 warm seconds on a pinned TaoMate, Light VAE and attention stack. The prefilled channel plays queued scenes; no viewer prompt-to-visible timing or full bill is reported. H3 derivative-weight rights have territorial limits. Read the setup and test boundary.

TaoMate-H3: window timing is not a stream bill

Project reports: a two-H200-class FP8 median of 4.74 seconds to generate a 4.958-second window on one server session. Viewer delay, synchronized audio on main, accepted quality and total cost remain unmeasured. Read the merge check.

fal Worlds compute-second rate

Both inspected fal model cards list $0.00125 per generated compute second: $4.50 for 3,600 charged seconds. Delivered duration, failures and full bill are unknown. Read the hosted-session check.

ShotStream: model throughput, not a delivered-hour rate

Paper reports: 15.95 generated FPS at 832×480 on one H200 for its Wan2.1-based story model. The released entrypoint saves a 16-FPS MP4 after inference. There is no source-timed receiver session, cold start, retry or H200 bill; no cost per generated or delivered hour follows. Read the check.

Veda-on-ComfyUI: whole clip, not just attention

Project says: on an RTX 5070 12 GB with ComfyUI 0.38.0, Windows 11, a 124-frame 1344×768 H3 T2VA job and eight-step Turbo LoRA, default attention took 342.1 seconds, Sage 230.9, and Veda 130.1. The advertised 7.1× compares attention time per step; default-to-Veda full-job gain was 2.63×. About 5.17 seconds of media still required 130.1 seconds. No accepted-output cost or live receiver test is available. Read the setup and caveat.

TaoMate-H3: first playable is later than first latent

Project says: on eight H20 96 GB GPUs, its ten-second 480×864 T2AV benchmark took 14.810 seconds of pure DiT time, 6.148 seconds to first final chunk latent, and 17.287 seconds to benchmark first playable video. The latter includes video VAE decode and H.264 publication but excludes internal audio preparation. No cold, repeated-block, receiver or billed-hour figure is available. Read the boundary.

Sol-H3: faster than playback before the final MP4

NVIDIA Research says: Sol-H3 took 1.653, 3.732 and 6.612 seconds for nominal 5-, 10- and 15-second 1344×768 T2VA audio-video outputs on eight B300s. These are warm medians through text encoding, four denoiser forwards and VAE decode. Loading, compilation, warmup and final MP4 encoding are outside the timer. The 11.04× five-second ratio also compares a 49-forward base profile with a four-forward FastH3 adapter; it is not an isolated runtime gain. No price, accepted-output rate, encoded-to-viewer delay or delivered-hour cost is supplied. Read the benchmark boundary.

Vidu S2-Editing: bill live seconds, not generated clips

Provider says: Editing bills one credit per on-live second. At Vidu’s $0.005 standard international credit price, 3,600 billed seconds imply $18 before tax, RTC, RTMP, storage, failures or delivery. The timer runs from successful live initialization to hangup; this is neither a measured bill nor a guaranteed viewer-delivered hour. Read the live-route boundary.

Four-step FastH3: denoiser speed is not the complete response

Project says: its current H3 recipe reports a warmed FastH3 Dense T2VA test on eight B300 GPUs: a 1344×768, 4.4-second clip took 11.7 and 11.8 seconds to return as a complete MP4. The diffusion stage took 2.37 and 2.36 seconds. Against the same-profile base H3 test, whole response improved about 2.2×, while the denoiser improved about 6.9×. This short result missed its own playback duration; it does not contradict the separate longer-clip profile below.

Read the timing and adapter conditions. No cost, first-frame, queue, delivered-stream, or repeated quality result is reported here.

Longer H3 output: no cost or wait figure yet

Project says: vLLM-Omni v0.30.0 returned a 30.667-second original-H3 Ref2VA clip from eight B300 GPUs. Its maintainer supplies no complete wall time or billing record. The 300-second continuation request bound cannot be turned into a cost per generated minute or delivered hour. Read the boundary.

ComfyUI H3 VAE: stage time is not job time

Comfy Org says: on an RTX 5090, H3 video VAE encode/decode at 1344×768 and 129 frames fell from 24.3 to 12.7 seconds after its v0.36.0 update. That is 11.6 seconds saved in VAE nodes. The report does not time denoising, queueing, retries, or delivery, so it does not support a new generated-hour or delivered-hour cost figure.

FastH3 V2 native canvas: one Radeon report, still far behind playback

Community report: one author reports FastH3 8-Step V2 on a Ryzen AI Max+ 395 / Radeon 8060S system with 128 GB unified memory, Windows, ROCm 10.0.0, and ComfyUI v0.36.0. The stated 768×1344, 243-frame job completed in 33 minutes 9 seconds end to end.

The arithmetic is 1,989 seconds divided by 243 ÷ 24. The author says VSA did not engage below 12,288 tokens, so a smaller 640×384 case did not gain from FastH3. This is one self-report, not a quality comparison, a transferable AMD result, a cold-start record, or a measure of FL2VA or live capacity.

New 5090 row: complete MP4, still slower than playback

Platform benchmark says: Sogni reports a pinned FastH3 Preview v1 image-to-video test on one RTX 5090 (32 GB). At 24 fps, its 362-frame output is 15.08 seconds long. The stated hot job-start-to-finished-MP4 result was 55 seconds at 480×640 and 126 seconds at 768×1024.

The arithmetic divides each stated wall time by 15.08 seconds. The source excludes queue and model-load time and says a 480p FastH3 canary after a worker restart took 160 seconds. It reports its own platform test and does not establish a general 5090 baseline, dollar cost, quality parity, retries, first-frame delivery, or stream capacity. This is four-step Preview v1, not the separate eight-step V2 Comfy package.

FastH3 on eight B300s: complete MP4 faster than playback

Project says: vLLM-Omni’s maintainer report used a warmed dedicated Dense/Data-Free FastH3 T2VA profile at 1344×768 / 24 fps with eight B300 GPUs. It returned complete 5.175-, 10.125-, and 15.083-second MP4s before their media duration ended.

The report excludes startup, compilation, and an excluded warmup. It says raw samples and logs are pending publication. It does not establish first frame, queue, concurrency, browser delivery, quality, cost, retries, uptime, FL2VA, or frame-carried continuity.

Two RTX 5090s: full H3 fits, but the output is offline

Project says: vLLM-Omni completed one 50-step T2VA job on two RTX 5090 GPUs: 124 frames at 1344×768 / 24 fps, about 5.17 seconds of H.264/AAC output, in 8 minutes 38 seconds client end to end. The recipe reports a full decode check and an approximate sampled peak of 22.6 GiB per GPU.

This is one T2VA run, not a warmed benchmark. It does not establish FL2VA/Ref2VA, first frame, queueing, cost, quality, delivery, or interactive response.

Older server row: full H3 on four B300s

Project says: vLLM-Omni’s current MiniMax H3 recipe reports a warmed FL2VA workload on four B300 GPUs: 209 frames at 1248×768, 50 steps, and 86.964 seconds mean HTTP client latency. At H3’s fixed 24 fps, 209 frames are about 8.71 seconds of output.

The calculation is 86.964 seconds divided by 209 ÷ 24 fps. The source excluded one warmup and averaged three requests. It does not include a rental price, cold start, endpoint queue, client rendering, transport, retries, quality review, or a viewer’s wait.

Do not compare the H3 server row with FastH3 by name alone

The vLLM route is a full-H3, 50-step FL2VA condition. Its current integration RFC says FastH3 requires a selected adapter fused at server start, and it is not switchable through the remote ComfyUI LoRA control. A few-step FastH3 result needs its own task, artifact, attention path, quality, hardware, warm state, and time boundary before it belongs in the same table.

Consumer channel: source-stated 22.1 fps, named conditions

Community project says: its FastH3 Live v1.2.0 route sustains 22.1 fps on one RTX 5090 (32 GB) under Windows 11 with ComfyUI. Its named setup uses a dense four-step conversion, a 448×448 canvas, a prequeue, retimed playout, and a 721-scene library.

FastH3 local: published timing, not a live claim

Project says: FastVideo compared the same 832×480, 124-frame, four-step FastH3 clip at 24 fps on an M4 Max, one DGX Spark, and two DGX Sparks. A 124-frame clip is about 5.17 seconds of output. The table converts its stated times into a decision number.

Current priced route: H3 Max Director

Provider says: the current fal page lists $0.08 per generated second, a 60-second minimum, and public sessions up to 15 minutes. That makes the stated minimum $4.80 and a stated full 15-minute session $72 in generated-video spend.

Unknown: actual billed time, direction-to-change latency, delivery cost, moderation cost, and whether an audience sees a clean restart. This arithmetic is not an operating quote.

Use this record