Every H3 thread has the same two questions in the comments: will it run on my card, and how long does it take. The answers are scattered across twenty threads and they contradict each other, so I pulled the numbers into one place.
The short version: the spread on a single card model is bigger than the spread between card models. Someone reported 22 minutes on a 5090 and got four replies saying they get the same work done in 3 to 5 minutes. Both are true. The difference is optimisation, not hardware.
Reported render times
| Hardware | Reported | Settings noted |
|---|---|---|
| 5090 | ~22 min, 362 frames | 768×1024, 8 steps, Turbo sampler |
| 5090 | 10:16 | 896×1120, no turbo, 20 steps |
| 5090 | "under 5 min" | not specified |
| 5090 | 134 s for 15 s clip | no turbo |
| 5090 | ~11 min for a 1:30 video | 12GB-optimised workflow, duration 30 |
| 3090 | ~3 min | 6-step turbo LoRA + attention optimisation |
| 5070 Ti + 64 GB | 118 s cold, 93 s warm | 608×352, default workflow |
| 5070 Ti + 64 GB | 230 s | 0.4 MP (480×864) |
| 4070 + 64 GB | 167 s | 608×352, default workflow |
| 3080 12 GB + 32 GB | works, Ref2VID | not timed |
| 12 GB VRAM | ~14 min for 30 s output | purpose-built low-VRAM workflow |
| 5060 Ti 16 GB + 128 GB | OOM | default t2v template, fresh install |
Sources are the comment threads on the 5090/362-frame post, the 12 GB 30-second workflow post, and the "12gb vram + 64ram" post. All within the last 30 days.
What explains a 4× spread on the same GPU
Reading through the replies, the same four things come up:
- Turbo LoRA with reduced steps. Multiple people running 4 to 6 steps instead of 20. This is the single largest factor.
- SageAttention / attention backend. Named repeatedly as the fix when someone reports a slow time.
- SLA speedup. One reply attributes the 22-minute figure directly to not using it.
- Resolution and megapixels. 608×352 versus 896×1120 is a large difference in work, and people quote times without stating either.
If you are benchmarking, state resolution, frame count, step count, and whether Turbo LoRA is on. Half the numbers floating around are unusable because they omit these.
The VRAM floor is lower than people assume, with a caveat
12 GB runs it. There are working reports on a 3080 12 GB with 32 GB system RAM, and a 12 GB-targeted workflow that produces 30-second output in about 14 minutes.
The caveat is that system RAM matters as much as VRAM here. The working 12 GB reports pair with 32 to 64 GB of system RAM. And a 5060 Ti with 16 GB VRAM and 128 GB system RAM still hit OOM on the default t2v template, which tells you the default template is not tuned for mid-range cards regardless of how much memory you have. Use a low-VRAM workflow rather than the default one and the floor drops.
"30-second generation" is stitched
Worth knowing before you plan around it: the 30-second low-VRAM workflow uses comfyui-h3-multishot, which joins three 10-second clips. It works, and the seams are handled, but it is not a native 30-second single pass. If your use case needs one continuous take with no join, test it before committing.
On the "$0 API cost" framing
The 5090 post was titled with "$0 API cost" and the top reply was:
"0$ cost" after pressing Queue button. but before that, Hardware cost quite much
That is the honest framing. Local generation is free per clip and expensive up front. Which one wins depends entirely on volume, and the break-even is easy to compute for your own situation:
clips_to_break_even = hardware_cost / (seconds_per_clip × api_rate_per_second)
I did not want to guess at anyone's GPU price or electricity rate, so plug in your own. For the API side of that equation, published per-second rates for hosted H3 are $0.074/s at 768P and $0.119/s at 2K (model page — disclosure: I work on reAPI). A 10-second 768P clip is $0.74, so at that rate a $2,000 card pays for itself somewhere around 2,700 clips, before electricity and before your own waiting time.
That number moves a lot with resolution and clip length, which is why the formula is more useful than any single figure I could quote.
Where each one actually wins
Local wins on: high volume, iteration where you are regenerating constantly, anything involving custom LoRAs, no content restrictions, and not caring about a 3-to-20-minute wait per attempt.
Hosted wins on: low or bursty volume, needing output now rather than after a tuning session, not owning a card that clears the floor, and 2K where local memory pressure gets ugly.
Most people posting here are firmly in the first group, which is why the local workflows get the upvotes. But the threads asking "can my 3060 do this" are a different group, and for them the arithmetic often does not favour buying a card.
What would help
If you have timings, post them with resolution, frames, steps, and whether Turbo LoRA and SageAttention are on. Right now the useful data is spread across comment replies in threads whose titles are about something else, and half of it is missing the settings that would make it comparable.
Top comments (0)