MiniMax H3 Singularity LoRA · Image to Video

"Community-tuned" is doing a lot of work in this node's description, so let's be clear about what it means. Singularity LoRA is a fine-tuned variant of MiniMax H3 tuned by people who aren't MiniMax. It isn't a bigger model, a smarter model, or a fixed model - it's a different aesthetic baked into the weights. If you don't like the look, the extra ~25% per second you're paying over base H3 buys you nothing.

What it is

The image-to-video node for MiniMax H3 Singularity LoRA, part of the PopNow pack: cloud inference, no local model, from about 9.38 credits ($0.0938) per second, 3 to 15 second clips with native audio, and a resolution ladder that goes further than most of the H3 nodes - 480p, 540p, 768p and 1080p.

The name ends in LoRA because it accepts custom LoRA weights on top, but the Singularity tuning is already in there. You're stacking your style on someone else's.

How it works

Your opening image is uploaded once (PNG, or JPEG if it'd exceed 10 MiB) and cached for 12 hours under a content hash, so prompt iterations don't re-upload. PopNow renders the clip on its own hardware and the node returns a ComfyUI VIDEO. Everything - image, prompt, LoRA files - leaves your machine, which is the price of not owning a GPU that can run a 33B video model.

The seed field works like a normal ComfyUI seed, with one twist worth reading: its control-after-generate is set to fixed by default. Queue the same graph twice and you get the same result rather than a second bill. Switch it to randomize when you actually want a new take. That's deliberate on the pack's part - every run here is metered.

The inputs

• prompt - what happens after the opening frame: action, camera, sound. Describe the motion you want, not the picture you already provided.
• image_url - the opening frame. The clip keeps its aspect ratio, so there's no separate aspect widget on this node; crop your source instead.
• resolution - 480p, 540p, 768p, 1080p. Default is 480p, and the price steps up sharply with each rung. Draft at 480p, final at 1080p, and don't finalise anything you haven't seen animated.
• duration_seconds - 3 to 15.
• seed - as above.
• last_image_url (optional) - a closing frame to resolve to.
• loras (optional) - up to three, each with its own strength.

Outputs are video (wire to Save Video) and task id (for the API console).

Adding your own LoRA

Drop a PopNow LoRA node in, set source and scale, connect it to loras, and chain nodes to stack up to three. source takes a .safetensors in ComfyUI/models/loras (uploaded once, kept 12 hours), a Hugging Face repo id or page, a Civitai page, or a direct https:// link. Gated files need HF_TOKEN or CIVITAI_API_KEY in the environment - never in a widget, since widget values end up in every workflow you share.

Train your LoRA against H3, not against Wan or Flux. A video LoRA from a different base won't do anything useful here, and the model will happily accept it and produce mush rather than tell you.

Install

Manager → search PopNow → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart, then add a key from popnow.ai/api/console/keys via Settings > PopNow and restart once more. The pack declares no Python dependencies; it runs on ComfyUI's own comfy_api.

When to skip it

Compare one 480p clip here against the same prompt on base H3 and the H3 LoRA node before committing. Base H3 is cheaper per second and reaches 2K; H3 LoRA is the variant you pick when your weights matter. Singularity is for when the fine-tune's own style is the point - moodier motion, punchier sound design, whatever the community tuning pushed. That's a taste decision, and taste decisions are worth a 50-cent test rather than a 12-clip batch.

Errors to expect: PopNow 40201 is credits, PopNow 400 naming a field is an unsupported value, and a request_id in the error is the string support needs.