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.

SGF+: a 24-hour paper result, a four-minute browser default, and a slower 5090 file route

Project reports: a B200 Reactor service generates about 19 FPS for 16-FPS browser playback; prompt edits wait for the next chunk and buffered frames. An independent ComfyUI pack needs 56.6 warm seconds for a 20-second 5090 output. Neither establishes a 24-hour viewer session or full cost. Compare the three routes.

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.

Reactor ComfyUI: a hosted live window, not local FastH3-Omni

Provider says: Reactor’s nodes connect ComfyUI to its API. FastH3-Omni queues finished audio-video clips before playing them; its displayed $0.75/minute rate is not a delivered-hour bill. Check the route before choosing hardware.

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.

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.

Self-Forcing stepwise serving is still a proposed route

An open vLLM-Omni PR adds a text-only WebSocket path with prompt updates. Its one-A800 local stream emits frames in chunks but takes 10.57 seconds for 5.06 seconds of 16-fps output. The separate minWM framework has camera controls and a completed-MP4 quickstart; it is not this streaming client. Wait for a merged route and receiver timing before budgeting a show.

Latest check: H3 video decoding after the first latent

Authors report: a decoded frame 0.04 seconds after the first H3 video latent exists. Their 30-clip reconstruction test loses 0.63 dB PSNR versus the official decoder. Prompt-to-viewer delay and cost are not measured. Read the decoder check.

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.

TaoMate-H3: runtime and LoRA are separate routes

Project says: current T2AV runtime schedules prompts by five-second blocks and writes a final MP4. Its validated benchmark is on eight H20 96 GB GPUs. A converted three-step LoRA used in ComfyUI on one RTX 5090 is a separate first-person offline report; it does not carry the official runtime's clean-KV continuity or prove a live session. Check the exact task, adapter, GPU, complete artifact, quality and receiver before choosing either path. Read the dated check.

vLLM-Omni H3: longer finished clips, not a stream

Project says: v0.30.0 adds opt-in Ref2VA latent-tail continuation. Its maintainer reports one 30.667-second original-H3 output on eight B300 GPUs at 1344×768 and 24 fps. The source does not give a complete wall time, cost, repeated-run result, or measured audio-video join.

Continuation needs Ref2VA request mode with uncached denoising. The 300-second request guard is not proof that five minutes works on your hardware. Use this for a longer finished-clip test, not a live-channel claim. Read the dated check.

Local H3 VAE: faster encode/decode, not a faster whole job by the same ratio

Comfy Org says: its v0.36.0 H3 video VAE update cut one RTX 5090, 1344×768, 129-frame encode/decode round trip from 24.3 to 12.7 seconds. The default fused encoder needs no flag. The faster optional path uses --fast fp16_accumulation; fastest decoding uses the int8 H3 VAE file.

This is a VAE-node A/B. Before saying your workflow is faster, record its full cold and warm render time, model load, denoise, VAE stages, MP4 output, quality, and failures. The source does not give a full-render result. Read the dated check.

Dreamverse FastH3: frame carry is a mechanism, not a measured stream

Project says: FastVideo’s current Dreamverse README exposes an opt-in FastH3 Preview v1 profile. It uses four visible GPUs by default, creates 124-frame 768×1344 audio-video segments, and supplies a segment’s last frame as first-frame conditioning to the next.

That distinguishes this route from a queue of unrelated clips. It does not establish a clean visual handoff, a warm segment rate, queue margin, visible response time, recovery, cost, or a consumer-GPU path. Its default compile warm-up can take tens of minutes on a cold cache; record readiness separately from the first finished segment.

FastH3 8-Step V2: the task boundary is still the first test

Project says: FastVideo’s current 35B 8-Step V2 model card and schedule call the checkpoint text-to-audio-video only, say FL2VA and Ref2VA were not distilled, and name four B200 GPUs for its trained default. ComfyUI says: its current image-to-video template for the same named checkpoint supports T2VA and first/last-frame FL2VA, excluding Ref2VA. Its paired text-to-video template says T2VA only and again says FL2VA was not distilled.

Do not call this a reliable frame-to-frame transition workflow until a pinned run checks it. Record the model-card and template revision, checkpoint, task, grid, scheduler, component files, attention path, GPU, system RAM, cold/warm result, and whether the supplied first and last frames visibly anchor the output.

FastH3 V2 on a Radeon 8060S: native canvas is not a live route

Community report: one author names a Ryzen AI Max+ 395, Radeon 8060S graphics, 128 GB unified memory, Windows, ROCm 10.0.0, and ComfyUI v0.36.0. The author reports a 768×1344, 243-frame V2 job in 33 minutes 9 seconds—about 10.1 seconds of output at 24 fps. Treat this as a condition-rich local-render lead, not an AMD baseline or a real-time result.

The author says the VSA path stays dense below 12,288 tokens. Their smaller 640×384 case did not gain from FastH3; the larger native-canvas case did. Put canvas, frame grid, token count, attention backend, memory topology, and the timing boundary beside any eight-step headline.

FastH3 Preview v1 on a 5090: a named timing, not a V2 result

Platform benchmark says: Sogni’s stated FastH3 Preview v1 image-to-video test ran a 15.08-second, 24 fps clip on one RTX 5090 (32 GB) in 55 seconds at 480×640 and 126 seconds at 768×1024. The hot boundary starts at job start and ends at a complete MP4; queue and model loading are excluded.

Use this as a review-loop or prequeue lead. It is not a normal-speed stream, a time-to-first-frame measurement, or a consumer result you can transfer to FastH3 8-Step V2. Record model family, task, attention path, checkpoint, source still, resolution, frames, audio setting, GPU, system RAM, and cold/warm state before comparing it with another FastH3 claim.

Remote H3 mask editing: API before ComfyUI

Project says: vLLM-Omni merged H3 source-media and noise-mask editing into its video API. A mask value of 0 preserves source; 1 regenerates it. The dedicated ComfyUI editing node remained an open PR at the 4 October check. Use direct API requests only if you can supply the documented mask format; do not assume a portable ComfyUI workflow or live edits. Read the client check.

vLLM-Omni: the complete-MP4 timing excludes viewer delivery

Project says: vLLM-Omni has a current H3 recipe for an OpenAI-compatible service that covers T2VA, FL2VA, and Ref2VA. Its recipe documents ordered first-and-last-frame inputs at the server layer. Its current H3 API/ComfyUI RFC says the client remains narrower: one FL2VA frame, at most two inputs per reference modality, and some valid mixed references rejected. The UI can run separately from the H3 GPU service, but it is not a complete proxy for the server contract.

The current recipe also says one 50-step full-H3 T2VA request completed on two RTX 5090 GPUs: 124 1344×768 / 24 fps frames, about 5.17 seconds of output, in 8 minutes 38 seconds client end to end. Its two-GPU DLO route needs at least 200 GiB host RAM, with 384 GiB recommended. That is an offline capacity validation, not a FL2VA result or a live route.

A separate maintainer report says a warmed, dedicated FastH3 Dense/Data-Free T2VA profile returned a complete 10.125-second 1344×768 / 24 fps MP4 in 8.678–8.710 seconds on eight B300 GPUs. This is complete response, not time to first frame or streaming. It excludes startup, compilation, and an excluded warmup; its raw bundle is pending.

Keep it apart from the project’s full-H3 four-B300 FL2VA baseline, which returns about 8.71 seconds at 24 fps in 86.964 seconds. The FastH3 result does not establish FL2VA, a request-switchable adapter, a browser-visible response, or a continuous transition route.

H3 batching: step scheduling is not a queue-speed claim

Project says: H3 can admit and retire requests between denoise steps. In its named two-H100 test, four 672×384, 209-frame, 30-step requests submitted together took 174.8 seconds in ordinary request mode and 182.1 seconds with four-way co-batching. Mean latency also rose from 111.5 to 175.7 seconds.

This is one simultaneous-arrival benchmark. The source leaves staggered arrivals unmeasured. Do not transfer it to a 5090, another task, or a viewer-visible stream. Record the arrival pattern and first playable media before claiming a queue is responsive.

The two-GPU H3 profile starts with a selected task

Project says: the current two-RTX-5090 DLO command selects the FL2VA DiT; its task contract maps text-only T2VA and FL2VA to that DiT. The recipe recommends a 384 GiB host. Its separate 8 minutes 38 seconds validation is one T2VA returned-artifact run, not an FL2VA timing or Ref2VA result.

Sol-H3 is a pinned fast-clip route, not a ComfyUI install promise

Project says: its eight-B300 benchmark uses a four-forward FastH3 Preview adapter at 1344×768 and 24 fps. The Sol-H3 code README says B300 SM103 is validated; reachable SM90 and SM120 paths still need hardware checks. Record the adapter, task, attention path, precision and output encoding before comparing it with another H3 route. See the measured boundary.

The task and adapter are part of the server

Project says: H3 uses different T2VA/FL2VA and Ref2VA task paths. The current vLLM-Omni RFC says FastH3 works by fusing one selected adapter at server start; its remote ComfyUI LoRA control cannot switch FastH3 for a request. It says the current client supports one FL2VA frame, so a first/last-frame plan stops at the client before it reaches the server.

Record the checkpoint, task, adapter filename, server revision, GPU and host memory, client revision, outgoing request, request shape, cold/warm state, response boundary, and audio result. An endpoint label alone is not a workflow.

One-5090 FastH3 channel: a format-specific community route

Community project says: FastH3 Live runs an unattended local audio-video channel on one RTX 5090 (32 GB) with ComfyUI. Its current card claims 22.1 sustained fps at 448×448 through a dense four-step FastH3 conversion, a prequeue, retimed playback, and a large library of self-contained scenes.

That is not a local interactive-session recipe. The source frames are authored for 24 fps; the project retimes delivery to its sustainable pace. Its scenes do not retain a viewer’s current world state across the next clip. Treat it as a reproducibility lead for a particular AI-TV format.

FastH3 has named attention paths

Project says: FastVideo’s current FastH3 Preview example selects VSA or dense FA4 attention from the adapter payload. An adapter carrying the VSA compression gate must run VSA; an adapter without it defaults to dense. Keep that fact beside any FastH3 speed number.

The one-5090 community project says it chose a dense conversion because stock ComfyUI does not preserve its VSA gate tensors. Its result is therefore not a consumer version of Reactor’s VSA B200 deployment. Same model family is not the same workflow.

Comfy’s official H3 access route: a footprint is not a speed claim

Comfy says: its optimized H3 path prunes modulation weights, uses INT8 ConvRot, and reduces the smallest variants’ total memory footprint from 123.6 GB to 42.5 GB. It says dynamic VRAM offloading can run H3 on a GPU like an RTX 3060.

That is useful for a first local test. It does not say which task mode, resolution, duration, quality, system RAM, retry rate, or wall time fits that card.

Record this before calling any workflow fast

Capture client and inference hosts, exact checkpoint, adapter, attention path, quantisation, canvas, source frames, delivered fps, task mode, host RAM, GPU, version, seed, cold and warm wall time, API time boundary, queue state, failures, and observed quality. Without those fields, the timing is not comparable.