I Ran H3 on AutoDL: Similar Results, Just Slower

Last time I used the MiniMax H3 API for the Gwen morph video and got a finished product for 35 RMB. After writing that post, the money-saving thought crept back in: H3 is open source now, right? Why not rent my own GPU, set it up once, and generate freely afterwards?

So today I started using AutoDL. The API bills per frame generated, which keeps a quiet meter running in your head; running it myself, that tension loosens.

AutoDL is a GPU rental platform, billed by the hour, and it does not charge while the machine is shut down. I topped up ten yuan and, at the price of the instance I picked that day, rented one, installed the environment, pulled the model, and opened ComfyUI. The H3 weights are not small, so downloading and unpacking them took a while. Moving to another instance may mean installing parts of the environment again.

Ten yuan does not sound like much, but it actually runs. The model is still H3, and the base output looks similar to the API — same model underneath, but not identical: the API can output 2K, while the self-hosted H3-Base tops out at 768p because the 2K modules are not open source. Saving money comes with a small resolution discount.

ComfyUI is the real barrier

After all the tinkering, I realized the GPU is not the barrier. ComfyUI is.

Before generating anything, I had to connect model loading, images, parameters, and output on a canvas. Templates exist, but they rarely run untouched: a path, plugin, or memory limit is always waiting to send you back through the nodes.

The worst failure is “everything looks normal, but the result is wrong.” No red node, no useful log, just a blurry output that makes you take the workflow apart piece by piece. Getting it connected is oddly satisfying, like finishing a LEGO set, and it does teach me more about the model. It also turned one API request into an evening of research.

It is not plug-and-play. Once it works, though, the whole process is mine to change.

Once it finally runs, generation is genuinely slow.

The cloud API sits behind a cluster; my rented card queues for inference alone. For a video, I stared at the progress bar and felt it was more hesitant than I was. It is slow, but while it runs I am not watching a bill tick up, so I can let it take its time. What really wears on you is tuning: with the API, changing one line of the prompt and rerunning is no big deal; locally, every change means re-inferring, and the wait is noticeably longer. The base output is close to the API, but the patience for "one more try should do it" runs out much faster on a local setup.

This run was genuinely cheap, and the hassle and slowness were real too. The money I saved seems to have been spent on patience instead.

Similar results and low cost are enough to keep me tinkering. For batch generation, or whenever I want to change the full pipeline, running it myself is still worthwhile. I just have to carry the waiting time.

Today's ten yuan was enough to get the environment and workflow running. Next time I wait for a progress bar, I will line up something else to do.

August 22 update: now it can generate directly

Not long after I wrote this, AutoDL added a more direct route. I can now pick a ready-made H3 workflow on AutoDL.Art and submit a job. It has text-to-video, first-and-last-frame generation, multi-image reference, and image-plus-audio workflows. I do not have to rent an instance first, download the model, install plugins, or wire the nodes myself.

It is genuinely cheap, too. The page lists a limited-time rate of RMB 0.01 per second for both 480p and 768p. My three calls all succeeded, and the records show charges of RMB 0.05, 0.10, and 0.15. That is almost absurdly cheap.

It is not instant. All three records include queue and inference time; the longest shows 1 minute 2 seconds in the queue and 8 minutes 3 seconds for inference. Compared with the rented-GPU route above, though, the model, environment, plugins, and workflow are no longer my problem.

The hassle from the first half of this post has not disappeared. AutoDL has just wrapped it behind the service. I would still rent a GPU when I want to change the full pipeline. If I only want to generate a clip, calling a hosted workflow is much easier and, at the promotional rate, much cheaper.