At 0.3 yuan per 15 seconds, T8 packs the MiniMax H3 viral-video workflow into a Skill any agent can install
Key numbers at a glance
One line: this is not a model review. It splits viral-video cloning into Skill, provider, digital human and upscaling, and pushes the cost low enough to count in dimes.
TL;DR: what this video actually ships
T8star-Aix published this 9 minute 56 second video on November 19, and it is not a model review. It is a delivery: the process of cloning viral short videos with MiniMax H3 is now a Skill package you install into any agent, the compute options expand from a single entry point to four, and the digital-human workflow gets a new base model plus a rewritten prompt node set.
The whole point is to push the dirty work that comes after generation — switching providers, counting cost, fixing faces, upscaling — into one reusable chain. Three things stand out:
- Cost lands at 0.3 yuan per 15 seconds. He uses his own intro film "Extraordinary Love" as the example: a 60 second cut required 90 seconds of generated footage to leave room for rejected takes, cost roughly 10 yuan, and about 5 yuan after a discount coupon.
- The Skill became an installable artifact. Four viral-clone Skills existed before; all were updated this round, and the key change is multi-provider support, which turns "which platform do I run on" into the first question after installation.
- Saving money and improving quality are separate tracks. Savings come from switching providers; quality comes from one-step RelayVSR upscaling and moving the digital-human base model to Dashiva V5 (a phonetic rendering from the spoken track).
He also sets the context: H3.1 is close, and whether it ships open weights is up to the vendor, so he will publish less H3 content for now. This video is effectively a wrap-up engineering pass on the existing H3 workflow.
What he delivered: three pieces
1. The Skill: copy it to an agent and let it install itself
His usage note is one sentence: copy the new Skill into whatever agent you use and tell it to install.
The first step after installation is not a run, but a question — the agent asks what inputs it needs. It wants images, one or two depending on the Skill, then it asks which provider to use. That "ask first, then run" design is the real difference this round: previously the pipeline was hardwired to one platform, now the provider is a variable.
His own build was redone after GPT-6.2 shipped and went through three versions, with gains in both output quality and cost. After installation you can map a specific image to a specific character and hand over an API key directly.
2. Providers: four routes, chosen by speed, price and patience
He splits compute into four categories with a plain selection logic:
- Local: free, one job at a time, and a one minute video can take hours. Fine for overnight batch work, painful when you need to revise during the day.
- RH (RunningHub): the previously published workflow still works, slightly more expensive, but ready to use.
- A low-cost online provider: the cheap entry he pushes here, 0.3 yuan per 15 second H3 clip, with low concurrency.
- Another online provider: newly added watermark-free video replacement, usable in the browser or through a direct API call.
(Apart from RH, the two other names are ambiguous in the spoken audio, so this article does not guess at spellings. Check the video description and the author's repository before buying.)
| Provider | Cost | Speed / concurrency | Best for |
|---|---|---|---|
| Local | free | one job at a time, hours per minute of video | overnight batch work |
| RH (RunningHub) | higher | existing workflow | people already on ComfyUI |
| Low-cost online | 0.3 yuan / 15s | low concurrency (a bit over 100 jobs) | bulk short video and short drama |
| Second online option | usage-based | browser and API | watermark-free replacement inside a pipeline |
He is candid about capacity: overall concurrency is somewhere above one hundred jobs, a bit more if H is included, and that is still not enough for large volumes of short drama production.
3. Digital human: new base model, rewritten prompt nodes
The digital-human pipeline is derived from his earlier "civil and martial" technique, but the changes are substantial:
- The base model moves to Dashiva V5 (phonetic), which he prefers for better identity consistency, better skin rendering, and more stable handling of wider motions.
- Generation runs on the Flex version of H3, because a digital human does not need much motion, which lets him drop a pile of acceleration LoRAs.
- The two 600 EMA entries are optional; he sets them very low and gives them only a little weight.
- An old AL model stays in the chain for one job: removing the greasy look. That step is a large part of why the output stops reading as AI.
- The prompt node was rewritten: beyond the original performance reconstruction, it adds several emotion tiers, colour styles (a Wong Kar-wai look, a "bright" look and others) and director styles.
His own evidence is anecdotal but telling: commenters started asking why the intro looks like a real person.
4. RelayVSR: one-step upscaling, and do not max the weight
The last segment adds RelayVSR. After the official update, you bring the base install up to date, download two models, and get a result in one step, quickly.
His only operational advice is specific: do not set the weight to 1; 0.5 to 0.6 is enough. Push it higher and faces go plastic, especially in close-ups, and a slightly soft image reads more like a real person. He adds that 5-series users can go a bit higher, without spelling out which cards or model versions he means.
That matches the official positioning. RelayVSR generates references only at sparse keyframes with a large model, then a lightweight network reconstructs every frame, which is why it is fast and why detail strength has to be balanced through parameters and source quality.
What the official material confirms
The RelayVSR project page (arXiv 2609.37850) reports engineering numbers rather than impressions: at 1080p on a single A100 80GB with a 15-frame keyframe interval and dual-endpoint conditioning, it reaches 29.29 FPS, 13.82 GB peak GPU memory and 0.327 s first-frame model latency, against 7.80 FPS, 24.447 GB and 2.83 s for FlashVSR-Tiny.
The repository (Apache-2.0) carries details that affect real use:
- It does 4× only. Output resolution is exactly four times the input, with no arbitrary scale factor.
- It does not copy audio. You mux the track back yourself.
- Weights are two files,
wan.safetensorsandflash.safetensors, each carrying its own configuration, with no runtime adapter merge. --window-size 16means a reference endpoint is generated every 15 frames; raise it to let the small model cover more intermediate frames.
So T8's "one step and fast" holds up under this architecture, and his "weight 0.5" is really a validated point on the curve between added detail and plastic skin.
Who it fits, and who should wait
Good fit:
- Teams making short dramas or bulk short video who are blocked by per-clip compute price rather than model capability.
- People with an existing ComfyUI or RH workflow who want to run the same graph on a cheaper provider.
- Anyone producing digital-human talking footage who cares about looking real rather than moving a lot.
Not a good fit:
- Anyone needing large concurrency. By his own account, a bit over one hundred jobs is not enough.
- Anyone needing arbitrary upscale factors or preserved audio tracks; RelayVSR does neither.
- Anyone waiting for H3.1. Whether it is open weights is up to the vendor, and rebuilding the workflow now may expire quickly.
- Anyone who needs the exact provider names to place an order: the spoken names are ambiguous, so verify them against the video description and repository.
How to use it: three steps
- Copy the Skill into your usual agent, install it, and let it list the inputs it needs.
- Supply one or two reference images as requested, then choose a provider (local, RH, a low-cost online option, or a browser/API option).
- Run the finished cut through RelayVSR at a weight of 0.5 to start, and mux the audio back separately.
Pitfalls
- Do not push the upscale weight to 1. Close-up faces turn plastic. Start at 0.5–0.6, check one clip, then adjust.
- Do not price a job from the final cut length. A 60 second cut consumed 90 seconds of generation; the rejection rate belongs in the budget.
- Do the math before running locally. Hours per minute of video makes revision brutal.
- Let the Skill ask its questions. Do not dump a prompt immediately; provider and image count are inputs it needs.
- Do not expect LoRA stacking to add motion. This pipeline assumes small motion and wins on expression and detail.
Next steps
- First, run one 15 second shortest-path job to confirm the provider works before planning volume.
- Second, run the same source clip on the cheap provider and on RH, and compare rejection rate, not just unit price.
- Third, put the rejection rate into the budget sheet instead of multiplying final runtime by unit price.
The short version
The value here is not the model. It is turning viral-video cloning from a craft into an engineering artifact you can package, re-point at a different provider, and install into an agent — with a cost line low enough to count in dimes.
> This article is compiled from a local speech-to-text transcript of the video, which has no subtitle track, and technical details are cross-checked against the official RelayVSR project page and repository. Model names taken from the spoken audio are phonetic; confirm spelling and version numbers against the official release.